crypto/evp/e_aes_cbc_hmac_sha256.c

   1 /* ====================================================================
   2  * Copyright (c) 2011-2013 The OpenSSL Project.  All rights reserved.
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  *
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  *
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in
  13  *    the documentation and/or other materials provided with the
  14  *    distribution.
  15  *
  16  * 3. All advertising materials mentioning features or use of this
  17  *    software must display the following acknowledgment:
  18  *    "This product includes software developed by the OpenSSL Project
  19  *    for use in the OpenSSL Toolkit. (http://www.OpenSSL.org/)"
  20  *
  21  * 4. The names "OpenSSL Toolkit" and "OpenSSL Project" must not be used to
  22  *    endorse or promote products derived from this software without
  23  *    prior written permission. For written permission, please contact
  24  *    licensing@OpenSSL.org.
  25  *
  26  * 5. Products derived from this software may not be called "OpenSSL"
  27  *    nor may "OpenSSL" appear in their names without prior written
  28  *    permission of the OpenSSL Project.
  29  *
  30  * 6. Redistributions of any form whatsoever must retain the following
  31  *    acknowledgment:
  32  *    "This product includes software developed by the OpenSSL Project
  33  *    for use in the OpenSSL Toolkit (http://www.OpenSSL.org/)"
  34  *
  35  * THIS SOFTWARE IS PROVIDED BY THE OpenSSL PROJECT ``AS IS'' AND ANY
  36  * EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  37  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  38  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE OpenSSL PROJECT OR
  39  * ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  40  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  41  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  42  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  43  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  44  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
  45  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
  46  * OF THE POSSIBILITY OF SUCH DAMAGE.
  47  * ====================================================================
  48  */
  49
  50 #include <openssl/opensslconf.h>
  51
  52 #include <stdio.h>
  53 #include <string.h>
  54
  55 #if !defined(OPENSSL_NO_AES) && !defined(OPENSSL_NO_SHA256)
  56
  57 #include <openssl/evp.h>
  58 #include <openssl/objects.h>
  59 #include <openssl/aes.h>
  60 #include <openssl/sha.h>
  61 #include <openssl/rand.h>
  62 #include "modes_lcl.h"
  63
  64 #ifndef EVP_CIPH_FLAG_AEAD_CIPHER
  65 #define EVP_CIPH_FLAG_AEAD_CIPHER       0x200000
  66 #define EVP_CTRL_AEAD_TLS1_AAD          0x16
  67 #define EVP_CTRL_AEAD_SET_MAC_KEY       0x17
  68 #endif
  69
  70 #if !defined(EVP_CIPH_FLAG_DEFAULT_ASN1)
  71 #define EVP_CIPH_FLAG_DEFAULT_ASN1 0
  72 #endif
  73
  74 #if !defined(EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK)
  75 #define EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK 0
  76 #endif
  77
  78 #define TLS1_1_VERSION 0x0302
  79
  80 typedef struct
  81     {
  82     AES_KEY             ks;
  83     SHA256_CTX          head,tail,md;
  84     size_t              payload_length; /* AAD length in decrypt case */
  85     union {
  86         unsigned int    tls_ver;
  87         unsigned char   tls_aad[16];    /* 13 used */
  88     } aux;
  89     } EVP_AES_HMAC_SHA256;
  90
  91 #define NO_PAYLOAD_LENGTH       ((size_t)-1)
  92
  93 #if     defined(AES_ASM) &&     ( \
  94         defined(__x86_64)       || defined(__x86_64__)  || \
  95         defined(_M_AMD64)       || defined(_M_X64)      || \
  96         defined(__INTEL__)      )
  97
  98 extern unsigned int OPENSSL_ia32cap_P[3];
  99 #define AESNI_CAPABLE   (1<<(57-32))
 100
 101 int aesni_set_encrypt_key(const unsigned char *userKey, int bits,
 102                               AES_KEY *key);
 103 int aesni_set_decrypt_key(const unsigned char *userKey, int bits,
 104                               AES_KEY *key);
 105
 106 void aesni_cbc_encrypt(const unsigned char *in,
 107                            unsigned char *out,
 108                            size_t length,
 109                            const AES_KEY *key,
 110                            unsigned char *ivec, int enc);
 111
 112 int aesni_cbc_sha256_enc (const void *inp, void *out, size_t blocks,
 113                 const AES_KEY *key, unsigned char iv[16],
 114                 SHA256_CTX *ctx,const void *in0);
 115
 116 #define data(ctx) ((EVP_AES_HMAC_SHA256 *)(ctx)->cipher_data)
 117
 118 static int aesni_cbc_hmac_sha256_init_key(EVP_CIPHER_CTX *ctx,
 119                         const unsigned char *inkey,
 120                         const unsigned char *iv, int enc)
 121         {
 122         EVP_AES_HMAC_SHA256 *key = data(ctx);
 123         int ret;
 124
 125         if (enc)
 126                 memset(&key->ks,0,sizeof(key->ks.rd_key)),
 127                 ret=aesni_set_encrypt_key(inkey,ctx->key_len*8,&key->ks);
 128         else
 129                 ret=aesni_set_decrypt_key(inkey,ctx->key_len*8,&key->ks);
 130
 131         SHA256_Init(&key->head);        /* handy when benchmarking */
 132         key->tail = key->head;
 133         key->md   = key->head;
 134
 135         key->payload_length = NO_PAYLOAD_LENGTH;
 136
 137         return ret<0?0:1;
 138         }
 139
 140 #define STITCHED_CALL
 141
 142 #if !defined(STITCHED_CALL)
 143 #define aes_off 0
 144 #endif
 145
 146 void sha256_block_data_order (void *c,const void *p,size_t len);
 147
 148 static void sha256_update(SHA256_CTX *c,const void *data,size_t len)
 149 {       const unsigned char *ptr = data;
 150         size_t res;
 151
 152         if ((res = c->num)) {
 153                 res = SHA256_CBLOCK-res;
 154                 if (len<res) res=len;
 155                 SHA256_Update (c,ptr,res);
 156                 ptr += res;
 157                 len -= res;
 158         }
 159
 160         res = len % SHA256_CBLOCK;
 161         len -= res;
 162
 163         if (len) {
 164                 sha256_block_data_order(c,ptr,len/SHA256_CBLOCK);
 165
 166                 ptr += len;
 167                 c->Nh += len>>29;
 168                 c->Nl += len<<=3;
 169                 if (c->Nl<(unsigned int)len) c->Nh++;
 170         }
 171
 172         if (res)
 173                 SHA256_Update(c,ptr,res);
 174 }
 175
 176 #ifdef SHA256_Update
 177 #undef SHA256_Update
 178 #endif
 179 #define SHA256_Update sha256_update
 180
 181 #if !defined(OPENSSL_NO_MULTIBLOCK) && EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK
 182
 183 typedef struct { unsigned int A[8],B[8],C[8],D[8],E[8],F[8],G[8],H[8]; } SHA256_MB_CTX;
 184 typedef struct { const unsigned char *ptr; int blocks;  } HASH_DESC;
 185
 186 void sha256_multi_block(SHA256_MB_CTX *,const HASH_DESC *,int);
 187
 188 typedef struct { const unsigned char *inp; unsigned char *out;
 189                  int blocks; u64 iv[2]; } CIPH_DESC;
 190
 191 void aesni_multi_cbc_encrypt(CIPH_DESC *,void *,int);
 192
 193 static size_t tls1_1_multi_block_encrypt(EVP_AES_HMAC_SHA256 *key,
 194         unsigned char *out, const unsigned char *inp, size_t inp_len,
 195         int n4x)        /* n4x is 1 or 2 */
 196 {
 197         HASH_DESC       hash_d[8], edges[8];
 198         CIPH_DESC       ciph_d[8];
 199         unsigned char   storage[sizeof(SHA256_MB_CTX)+32];
 200         union { u64     q[16];
 201                 u32     d[32];
 202                 u8      c[128]; } blocks[8];
 203         SHA256_MB_CTX   *ctx;
 204         unsigned int    frag, last, packlen, i, x4=4*n4x, minblocks, processed=0;
 205         size_t          ret = 0;
 206         u8              *IVs;
 207 #if defined(BSWAP8)
 208         u64             seqnum;
 209 #endif
 210
 211         /* ask for IVs in bulk */
 212         if (RAND_bytes((IVs=blocks[0].c),16*x4)<=0)
 213                 return 0;
 214
 215         /* align */
 216         ctx = (SHA256_MB_CTX *)(storage+32-((size_t)storage%32));
 217
 218         frag = (unsigned int)inp_len>>(1+n4x);
 219         last = (unsigned int)inp_len+frag-(frag<<(1+n4x));
 220         if (last>frag && ((last+13+9)%64)<(x4-1)) {
 221                 frag++;
 222                 last -= x4-1;
 223         }
 224
 225         packlen = 5+16+((frag+32+16)&-16);
 226
 227         /* populate descriptors with pointers and IVs */
 228         hash_d[0].ptr = inp;
 229         ciph_d[0].inp = inp;
 230         /* 5+16 is place for header and explicit IV */
 231         ciph_d[0].out = out+5+16;
 232         memcpy(ciph_d[0].out-16,IVs,16);
 233         memcpy(ciph_d[0].iv,IVs,16);    IVs += 16;
 234
 235         for (i=1;i<x4;i++) {
 236                 ciph_d[i].inp = hash_d[i].ptr = hash_d[i-1].ptr+frag;
 237                 ciph_d[i].out = ciph_d[i-1].out+packlen;
 238                 memcpy(ciph_d[i].out-16,IVs,16);
 239                 memcpy(ciph_d[i].iv,IVs,16);    IVs+=16;
 240         }
 241
 242 #if defined(BSWAP8)
 243         memcpy(blocks[0].c,key->md.data,8);
 244         seqnum = BSWAP8(blocks[0].q[0]);
 245 #endif
 246         for (i=0;i<x4;i++) {
 247                 unsigned int len = (i==(x4-1)?last:frag);
 248 #if !defined(BSWAP8)
 249                 unsigned int carry, j;
 250 #endif
 251
 252                 ctx->A[i] = key->md.h[0];
 253                 ctx->B[i] = key->md.h[1];
 254                 ctx->C[i] = key->md.h[2];
 255                 ctx->D[i] = key->md.h[3];
 256                 ctx->E[i] = key->md.h[4];
 257                 ctx->F[i] = key->md.h[5];
 258                 ctx->G[i] = key->md.h[6];
 259                 ctx->H[i] = key->md.h[7];
 260
 261                 /* fix seqnum */
 262 #if defined(BSWAP8)
 263                 blocks[i].q[0] = BSWAP8(seqnum+i);
 264 #else
 265                 for (carry=i,j=8;j--;) {
 266                         blocks[i].c[j] = ((u8*)key->md.data)[j]+carry;
 267                         carry = (blocks[i].c[j]-carry)>>(sizeof(carry)*8-1);
 268                 }
 269 #endif
 270                 blocks[i].c[8] = ((u8*)key->md.data)[8];
 271                 blocks[i].c[9] = ((u8*)key->md.data)[9];
 272                 blocks[i].c[10] = ((u8*)key->md.data)[10];
 273                 /* fix length */
 274                 blocks[i].c[11] = (u8)(len>>8);
 275                 blocks[i].c[12] = (u8)(len);
 276
 277                 memcpy(blocks[i].c+13,hash_d[i].ptr,64-13);
 278                 hash_d[i].ptr += 64-13;
 279                 hash_d[i].blocks = (len-(64-13))/64;
 280
 281                 edges[i].ptr = blocks[i].c;
 282                 edges[i].blocks = 1;
 283         }
 284
 285         /* hash 13-byte headers and first 64-13 bytes of inputs */
 286         sha256_multi_block(ctx,edges,n4x);
 287         /* hash bulk inputs */
 288 #define MAXCHUNKSIZE    2048
 289 #if     MAXCHUNKSIZE%64
 290 #error  "MAXCHUNKSIZE is not divisible by 64"
 291 #elif   MAXCHUNKSIZE
 292         /* goal is to minimize pressure on L1 cache by moving
 293          * in shorter steps, so that hashed data is still in
 294          * the cache by the time we encrypt it */
 295         minblocks = ((frag<=last ? frag : last)-(64-13))/64;
 296         if (minblocks>MAXCHUNKSIZE/64) {
 297                 for (i=0;i<x4;i++) {
 298                         edges[i].ptr     = hash_d[i].ptr;
 299                         edges[i].blocks  = MAXCHUNKSIZE/64;
 300                         ciph_d[i].blocks = MAXCHUNKSIZE/16;
 301                 }
 302                 do {
 303                         sha256_multi_block(ctx,edges,n4x);
 304                         aesni_multi_cbc_encrypt(ciph_d,&key->ks,n4x);
 305
 306                         for (i=0;i<x4;i++) {
 307                                 edges[i].ptr     = hash_d[i].ptr += MAXCHUNKSIZE;
 308                                 hash_d[i].blocks -= MAXCHUNKSIZE/64;
 309                                 edges[i].blocks  = MAXCHUNKSIZE/64;
 310                                 ciph_d[i].inp    += MAXCHUNKSIZE;
 311                                 ciph_d[i].out    += MAXCHUNKSIZE;
 312                                 ciph_d[i].blocks = MAXCHUNKSIZE/16;
 313                                 memcpy(ciph_d[i].iv,ciph_d[i].out-16,16);
 314                         }
 315                         processed += MAXCHUNKSIZE;
 316                         minblocks -= MAXCHUNKSIZE/64;
 317                 } while (minblocks>MAXCHUNKSIZE/64);
 318         }
 319 #endif
 320 #undef  MAXCHUNKSIZE
 321         sha256_multi_block(ctx,hash_d,n4x);
 322
 323         memset(blocks,0,sizeof(blocks));
 324         for (i=0;i<x4;i++) {
 325                 unsigned int            len = (i==(x4-1)?last:frag),
 326                                         off = hash_d[i].blocks*64;
 327                 const unsigned char    *ptr = hash_d[i].ptr+off;
 328
 329                 off = (len-processed)-(64-13)-off;      /* remainder actually */
 330                 memcpy(blocks[i].c,ptr,off);
 331                 blocks[i].c[off]=0x80;
 332                 len += 64+13;           /* 64 is HMAC header */
 333                 len *= 8;               /* convert to bits */
 334                 if (off<(64-8)) {
 335 #ifdef BSWAP4
 336                         blocks[i].d[15] = BSWAP4(len);
 337 #else
 338                         PUTU32(blocks[i].c+60,len);
 339 #endif
 340                         edges[i].blocks = 1;
 341                 } else {
 342 #ifdef BSWAP4
 343                         blocks[i].d[31] = BSWAP4(len);
 344 #else
 345                         PUTU32(blocks[i].c+124,len);
 346 #endif
 347                         edges[i].blocks = 2;
 348                 }
 349                 edges[i].ptr = blocks[i].c;
 350         }
 351
 352         /* hash input tails and finalize */
 353         sha256_multi_block(ctx,edges,n4x);
 354
 355         memset(blocks,0,sizeof(blocks));
 356         for (i=0;i<x4;i++) {
 357 #ifdef BSWAP4
 358                 blocks[i].d[0] = BSWAP4(ctx->A[i]);     ctx->A[i] = key->tail.h[0];
 359                 blocks[i].d[1] = BSWAP4(ctx->B[i]);     ctx->B[i] = key->tail.h[1];
 360                 blocks[i].d[2] = BSWAP4(ctx->C[i]);     ctx->C[i] = key->tail.h[2];
 361                 blocks[i].d[3] = BSWAP4(ctx->D[i]);     ctx->D[i] = key->tail.h[3];
 362                 blocks[i].d[4] = BSWAP4(ctx->E[i]);     ctx->E[i] = key->tail.h[4];
 363                 blocks[i].d[5] = BSWAP4(ctx->F[i]);     ctx->F[i] = key->tail.h[5];
 364                 blocks[i].d[6] = BSWAP4(ctx->G[i]);     ctx->G[i] = key->tail.h[6];
 365                 blocks[i].d[7] = BSWAP4(ctx->H[i]);     ctx->H[i] = key->tail.h[7];
 366                 blocks[i].c[32] = 0x80;
 367                 blocks[i].d[15] = BSWAP4((64+32)*8);
 368 #else
 369                 PUTU32(blocks[i].c+0,ctx->A[i]);        ctx->A[i] = key->tail.h[0];
 370                 PUTU32(blocks[i].c+4,ctx->B[i]);        ctx->B[i] = key->tail.h[1];
 371                 PUTU32(blocks[i].c+8,ctx->C[i]);        ctx->C[i] = key->tail.h[2];
 372                 PUTU32(blocks[i].c+12,ctx->D[i]);       ctx->D[i] = key->tail.h[3];
 373                 PUTU32(blocks[i].c+16,ctx->E[i]);       ctx->E[i] = key->tail.h[4];
 374                 PUTU32(blocks[i].c+20,ctx->F[i]);       ctx->F[i] = key->tail.h[5];
 375                 PUTU32(blocks[i].c+24,ctx->G[i]);       ctx->G[i] = key->tail.h[6];
 376                 PUTU32(blocks[i].c+28,ctx->H[i]);       ctx->H[i] = key->tail.h[7];
 377                 blocks[i].c[32] = 0x80;
 378                 PUTU32(blocks[i].c+60,(64+32)*8);
 379 #endif
 380                 edges[i].ptr = blocks[i].c;
 381                 edges[i].blocks = 1;
 382         }
 383
 384         /* finalize MACs */
 385         sha256_multi_block(ctx,edges,n4x);
 386
 387         for (i=0;i<x4;i++) {
 388                 unsigned int len = (i==(x4-1)?last:frag), pad, j;
 389                 unsigned char *out0 = out;
 390
 391                 memcpy(ciph_d[i].out,ciph_d[i].inp,len-processed);
 392                 ciph_d[i].inp = ciph_d[i].out;
 393
 394                 out += 5+16+len;
 395
 396                 /* write MAC */
 397                 PUTU32(out+0,ctx->A[i]);
 398                 PUTU32(out+4,ctx->B[i]);
 399                 PUTU32(out+8,ctx->C[i]);
 400                 PUTU32(out+12,ctx->D[i]);
 401                 PUTU32(out+16,ctx->E[i]);
 402                 PUTU32(out+20,ctx->F[i]);
 403                 PUTU32(out+24,ctx->G[i]);
 404                 PUTU32(out+28,ctx->H[i]);
 405                 out += 32;
 406                 len += 32;
 407
 408                 /* pad */
 409                 pad = 15-len%16;
 410                 for (j=0;j<=pad;j++) *(out++) = pad;
 411                 len += pad+1;
 412
 413                 ciph_d[i].blocks = (len-processed)/16;
 414                 len += 16;      /* account for explicit iv */
 415
 416                 /* arrange header */
 417                 out0[0] = ((u8*)key->md.data)[8];
 418                 out0[1] = ((u8*)key->md.data)[9];
 419                 out0[2] = ((u8*)key->md.data)[10];
 420                 out0[3] = (u8)(len>>8);
 421                 out0[4] = (u8)(len);
 422
 423                 ret += len+5;
 424                 inp += frag;
 425         }
 426
 427         aesni_multi_cbc_encrypt(ciph_d,&key->ks,n4x);
 428
 429         OPENSSL_cleanse(blocks,sizeof(blocks));
 430         OPENSSL_cleanse(ctx,sizeof(*ctx));
 431
 432         return ret;
 433 }
 434 #endif
 435
 436 static int aesni_cbc_hmac_sha256_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
 437                       const unsigned char *in, size_t len)
 438         {
 439         EVP_AES_HMAC_SHA256 *key = data(ctx);
 440         unsigned int l;
 441         size_t  plen = key->payload_length,
 442                 iv = 0,         /* explicit IV in TLS 1.1 and later */
 443                 sha_off = 0;
 444 #if defined(STITCHED_CALL)
 445         size_t  aes_off = 0,
 446                 blocks;
 447
 448         sha_off = SHA256_CBLOCK-key->md.num;
 449 #endif
 450
 451         key->payload_length = NO_PAYLOAD_LENGTH;
 452
 453         if (len%AES_BLOCK_SIZE) return 0;
 454
 455         if (ctx->encrypt) {
 456                 if (plen==NO_PAYLOAD_LENGTH)
 457                         plen = len;
 458                 else if (len!=((plen+SHA256_DIGEST_LENGTH+AES_BLOCK_SIZE)&-AES_BLOCK_SIZE))
 459                         return 0;
 460                 else if (key->aux.tls_ver >= TLS1_1_VERSION)
 461                         iv = AES_BLOCK_SIZE;
 462
 463 #if defined(STITCHED_CALL)
 464                 if (OPENSSL_ia32cap_P[1]&(1<<(60-32)) && /* AVX? */
 465                     plen>(sha_off+iv) &&
 466                     (blocks=(plen-(sha_off+iv))/SHA256_CBLOCK)) {
 467                         SHA256_Update(&key->md,in+iv,sha_off);
 468
 469                         (void)aesni_cbc_sha256_enc(in,out,blocks,&key->ks,
 470                                 ctx->iv,&key->md,in+iv+sha_off);
 471                         blocks *= SHA256_CBLOCK;
 472                         aes_off += blocks;
 473                         sha_off += blocks;
 474                         key->md.Nh += blocks>>29;
 475                         key->md.Nl += blocks<<=3;
 476                         if (key->md.Nl<(unsigned int)blocks) key->md.Nh++;
 477                 } else {
 478                         sha_off = 0;
 479                 }
 480 #endif
 481                 sha_off += iv;
 482                 SHA256_Update(&key->md,in+sha_off,plen-sha_off);
 483
 484                 if (plen!=len)  {       /* "TLS" mode of operation */
 485                         if (in!=out)
 486                                 memcpy(out+aes_off,in+aes_off,plen-aes_off);
 487
 488                         /* calculate HMAC and append it to payload */
 489                         SHA256_Final(out+plen,&key->md);
 490                         key->md = key->tail;
 491                         SHA256_Update(&key->md,out+plen,SHA256_DIGEST_LENGTH);
 492                         SHA256_Final(out+plen,&key->md);
 493
 494                         /* pad the payload|hmac */
 495                         plen += SHA256_DIGEST_LENGTH;
 496                         for (l=len-plen-1;plen<len;plen++) out[plen]=l;
 497                         /* encrypt HMAC|padding at once */
 498                         aesni_cbc_encrypt(out+aes_off,out+aes_off,len-aes_off,
 499                                         &key->ks,ctx->iv,1);
 500                 } else {
 501                         aesni_cbc_encrypt(in+aes_off,out+aes_off,len-aes_off,
 502                                         &key->ks,ctx->iv,1);
 503                 }
 504         } else {
 505                 union { unsigned int  u[SHA256_DIGEST_LENGTH/sizeof(unsigned int)];
 506                         unsigned char c[64+SHA256_DIGEST_LENGTH]; } mac, *pmac;
 507
 508                 /* arrange cache line alignment */
 509                 pmac = (void *)(((size_t)mac.c+63)&((size_t)0-64));
 510
 511                 /* decrypt HMAC|padding at once */
 512                 aesni_cbc_encrypt(in,out,len,
 513                                 &key->ks,ctx->iv,0);
 514
 515                 if (plen != NO_PAYLOAD_LENGTH) {        /* "TLS" mode of operation */
 516                         size_t inp_len, mask, j, i;
 517                         unsigned int res, maxpad, pad, bitlen;
 518                         int ret = 1;
 519                         union { unsigned int  u[SHA_LBLOCK];
 520                                 unsigned char c[SHA256_CBLOCK]; }
 521                                 *data = (void *)key->md.data;
 522
 523                         if ((key->aux.tls_aad[plen-4]<<8|key->aux.tls_aad[plen-3])
 524                             >= TLS1_1_VERSION)
 525                                 iv = AES_BLOCK_SIZE;
 526
 527                         if (len<(iv+SHA256_DIGEST_LENGTH+1))
 528                                 return 0;
 529
 530                         /* omit explicit iv */
 531                         out += iv;
 532                         len -= iv;
 533
 534                         /* figure out payload length */
 535                         pad = out[len-1];
 536                         maxpad = len-(SHA256_DIGEST_LENGTH+1);
 537                         maxpad |= (255-maxpad)>>(sizeof(maxpad)*8-8);
 538                         maxpad &= 255;
 539
 540                         inp_len = len - (SHA256_DIGEST_LENGTH+pad+1);
 541                         mask = (0-((inp_len-len)>>(sizeof(inp_len)*8-1)));
 542                         inp_len &= mask;
 543                         ret &= (int)mask;
 544
 545                         key->aux.tls_aad[plen-2] = inp_len>>8;
 546                         key->aux.tls_aad[plen-1] = inp_len;
 547
 548                         /* calculate HMAC */
 549                         key->md = key->head;
 550                         SHA256_Update(&key->md,key->aux.tls_aad,plen);
 551
 552 #if 1
 553                         len -= SHA256_DIGEST_LENGTH;            /* amend mac */
 554                         if (len>=(256+SHA256_CBLOCK)) {
 555                                 j = (len-(256+SHA256_CBLOCK))&(0-SHA256_CBLOCK);
 556                                 j += SHA256_CBLOCK-key->md.num;
 557                                 SHA256_Update(&key->md,out,j);
 558                                 out += j;
 559                                 len -= j;
 560                                 inp_len -= j;
 561                         }
 562
 563                         /* but pretend as if we hashed padded payload */
 564                         bitlen = key->md.Nl+(inp_len<<3);       /* at most 18 bits */
 565 #ifdef BSWAP4
 566                         bitlen = BSWAP4(bitlen);
 567 #else
 568                         mac.c[0] = 0;
 569                         mac.c[1] = (unsigned char)(bitlen>>16);
 570                         mac.c[2] = (unsigned char)(bitlen>>8);
 571                         mac.c[3] = (unsigned char)bitlen;
 572                         bitlen = mac.u[0];
 573 #endif
 574
 575                         pmac->u[0]=0;
 576                         pmac->u[1]=0;
 577                         pmac->u[2]=0;
 578                         pmac->u[3]=0;
 579                         pmac->u[4]=0;
 580                         pmac->u[5]=0;
 581                         pmac->u[6]=0;
 582                         pmac->u[7]=0;
 583
 584                         for (res=key->md.num, j=0;j<len;j++) {
 585                                 size_t c = out[j];
 586                                 mask = (j-inp_len)>>(sizeof(j)*8-8);
 587                                 c &= mask;
 588                                 c |= 0x80&~mask&~((inp_len-j)>>(sizeof(j)*8-8));
 589                                 data->c[res++]=(unsigned char)c;
 590
 591                                 if (res!=SHA256_CBLOCK) continue;
 592
 593                                 /* j is not incremented yet */
 594                                 mask = 0-((inp_len+7-j)>>(sizeof(j)*8-1));
 595                                 data->u[SHA_LBLOCK-1] |= bitlen&mask;
 596                                 sha256_block_data_order(&key->md,data,1);
 597                                 mask &= 0-((j-inp_len-72)>>(sizeof(j)*8-1));
 598                                 pmac->u[0] |= key->md.h[0] & mask;
 599                                 pmac->u[1] |= key->md.h[1] & mask;
 600                                 pmac->u[2] |= key->md.h[2] & mask;
 601                                 pmac->u[3] |= key->md.h[3] & mask;
 602                                 pmac->u[4] |= key->md.h[4] & mask;
 603                                 pmac->u[5] |= key->md.h[5] & mask;
 604                                 pmac->u[6] |= key->md.h[6] & mask;
 605                                 pmac->u[7] |= key->md.h[7] & mask;
 606                                 res=0;
 607                         }
 608
 609                         for(i=res;i<SHA256_CBLOCK;i++,j++) data->c[i]=0;
 610
 611                         if (res>SHA256_CBLOCK-8) {
 612                                 mask = 0-((inp_len+8-j)>>(sizeof(j)*8-1));
 613                                 data->u[SHA_LBLOCK-1] |= bitlen&mask;
 614                                 sha256_block_data_order(&key->md,data,1);
 615                                 mask &= 0-((j-inp_len-73)>>(sizeof(j)*8-1));
 616                                 pmac->u[0] |= key->md.h[0] & mask;
 617                                 pmac->u[1] |= key->md.h[1] & mask;
 618                                 pmac->u[2] |= key->md.h[2] & mask;
 619                                 pmac->u[3] |= key->md.h[3] & mask;
 620                                 pmac->u[4] |= key->md.h[4] & mask;
 621                                 pmac->u[5] |= key->md.h[5] & mask;
 622                                 pmac->u[6] |= key->md.h[6] & mask;
 623                                 pmac->u[7] |= key->md.h[7] & mask;
 624
 625                                 memset(data,0,SHA256_CBLOCK);
 626                                 j+=64;
 627                         }
 628                         data->u[SHA_LBLOCK-1] = bitlen;
 629                         sha256_block_data_order(&key->md,data,1);
 630                         mask = 0-((j-inp_len-73)>>(sizeof(j)*8-1));
 631                         pmac->u[0] |= key->md.h[0] & mask;
 632                         pmac->u[1] |= key->md.h[1] & mask;
 633                         pmac->u[2] |= key->md.h[2] & mask;
 634                         pmac->u[3] |= key->md.h[3] & mask;
 635                         pmac->u[4] |= key->md.h[4] & mask;
 636                         pmac->u[5] |= key->md.h[5] & mask;
 637                         pmac->u[6] |= key->md.h[6] & mask;
 638                         pmac->u[7] |= key->md.h[7] & mask;
 639
 640 #ifdef BSWAP4
 641                         pmac->u[0] = BSWAP4(pmac->u[0]);
 642                         pmac->u[1] = BSWAP4(pmac->u[1]);
 643                         pmac->u[2] = BSWAP4(pmac->u[2]);
 644                         pmac->u[3] = BSWAP4(pmac->u[3]);
 645                         pmac->u[4] = BSWAP4(pmac->u[4]);
 646                         pmac->u[5] = BSWAP4(pmac->u[5]);
 647                         pmac->u[6] = BSWAP4(pmac->u[6]);
 648                         pmac->u[7] = BSWAP4(pmac->u[7]);
 649 #else
 650                         for (i=0;i<8;i++) {
 651                                 res = pmac->u[i];
 652                                 pmac->c[4*i+0]=(unsigned char)(res>>24);
 653                                 pmac->c[4*i+1]=(unsigned char)(res>>16);
 654                                 pmac->c[4*i+2]=(unsigned char)(res>>8);
 655                                 pmac->c[4*i+3]=(unsigned char)res;
 656                         }
 657 #endif
 658                         len += SHA256_DIGEST_LENGTH;
 659 #else
 660                         SHA256_Update(&key->md,out,inp_len);
 661                         res = key->md.num;
 662                         SHA256_Final(pmac->c,&key->md);
 663
 664                         {
 665                         unsigned int inp_blocks, pad_blocks;
 666
 667                         /* but pretend as if we hashed padded payload */
 668                         inp_blocks = 1+((SHA256_CBLOCK-9-res)>>(sizeof(res)*8-1));
 669                         res += (unsigned int)(len-inp_len);
 670                         pad_blocks = res / SHA256_CBLOCK;
 671                         res %= SHA256_CBLOCK;
 672                         pad_blocks += 1+((SHA256_CBLOCK-9-res)>>(sizeof(res)*8-1));
 673                         for (;inp_blocks<pad_blocks;inp_blocks++)
 674                                 sha1_block_data_order(&key->md,data,1);
 675                         }
 676 #endif
 677                         key->md = key->tail;
 678                         SHA256_Update(&key->md,pmac->c,SHA256_DIGEST_LENGTH);
 679                         SHA256_Final(pmac->c,&key->md);
 680
 681                         /* verify HMAC */
 682                         out += inp_len;
 683                         len -= inp_len;
 684 #if 1
 685                         {
 686                         unsigned char *p = out+len-1-maxpad-SHA256_DIGEST_LENGTH;
 687                         size_t off = out-p;
 688                         unsigned int c, cmask;
 689
 690                         maxpad += SHA256_DIGEST_LENGTH;
 691                         for (res=0,i=0,j=0;j<maxpad;j++) {
 692                                 c = p[j];
 693                                 cmask = ((int)(j-off-SHA256_DIGEST_LENGTH))>>(sizeof(int)*8-1);
 694                                 res |= (c^pad)&~cmask;  /* ... and padding */
 695                                 cmask &= ((int)(off-1-j))>>(sizeof(int)*8-1);
 696                                 res |= (c^pmac->c[i])&cmask;
 697                                 i += 1&cmask;
 698                         }
 699                         maxpad -= SHA256_DIGEST_LENGTH;
 700
 701                         res = 0-((0-res)>>(sizeof(res)*8-1));
 702                         ret &= (int)~res;
 703                         }
 704 #else
 705                         for (res=0,i=0;i<SHA256_DIGEST_LENGTH;i++)
 706                                 res |= out[i]^pmac->c[i];
 707                         res = 0-((0-res)>>(sizeof(res)*8-1));
 708                         ret &= (int)~res;
 709
 710                         /* verify padding */
 711                         pad = (pad&~res) | (maxpad&res);
 712                         out = out+len-1-pad;
 713                         for (res=0,i=0;i<pad;i++)
 714                                 res |= out[i]^pad;
 715
 716                         res = (0-res)>>(sizeof(res)*8-1);
 717                         ret &= (int)~res;
 718 #endif
 719                         return ret;
 720                 } else {
 721                         SHA256_Update(&key->md,out,len);
 722                 }
 723         }
 724
 725         return 1;
 726         }
 727
 728 static int aesni_cbc_hmac_sha256_ctrl(EVP_CIPHER_CTX *ctx, int type, int arg, void *ptr)
 729         {
 730         EVP_AES_HMAC_SHA256 *key = data(ctx);
 731         unsigned int u_arg = (unsigned int)arg;
 732
 733         switch (type)
 734                 {
 735         case EVP_CTRL_AEAD_SET_MAC_KEY:
 736                 {
 737                 unsigned int  i;
 738                 unsigned char hmac_key[64];
 739
 740                 memset (hmac_key,0,sizeof(hmac_key));
 741
 742                 if (arg < 0)
 743                         return -1;
 744
 745                 if (u_arg > sizeof(hmac_key)) {
 746                         SHA256_Init(&key->head);
 747                         SHA256_Update(&key->head,ptr,arg);
 748                         SHA256_Final(hmac_key,&key->head);
 749                 } else {
 750                         memcpy(hmac_key,ptr,arg);
 751                 }
 752
 753                 for (i=0;i<sizeof(hmac_key);i++)
 754                         hmac_key[i] ^= 0x36;            /* ipad */
 755                 SHA256_Init(&key->head);
 756                 SHA256_Update(&key->head,hmac_key,sizeof(hmac_key));
 757
 758                 for (i=0;i<sizeof(hmac_key);i++)
 759                         hmac_key[i] ^= 0x36^0x5c;       /* opad */
 760                 SHA256_Init(&key->tail);
 761                 SHA256_Update(&key->tail,hmac_key,sizeof(hmac_key));
 762
 763                 OPENSSL_cleanse(hmac_key,sizeof(hmac_key));
 764
 765                 return 1;
 766                 }
 767         case EVP_CTRL_AEAD_TLS1_AAD:
 768                 {
 769                 unsigned char *p=ptr;
 770                 unsigned int   len=p[arg-2]<<8|p[arg-1];
 771
 772                 if (ctx->encrypt)
 773                         {
 774                         key->payload_length = len;
 775                         if ((key->aux.tls_ver=p[arg-4]<<8|p[arg-3]) >= TLS1_1_VERSION) {
 776                                 len -= AES_BLOCK_SIZE;
 777                                 p[arg-2] = len>>8;
 778                                 p[arg-1] = len;
 779                         }
 780                         key->md = key->head;
 781                         SHA256_Update(&key->md,p,arg);
 782
 783                         return (int)(((len+SHA256_DIGEST_LENGTH+AES_BLOCK_SIZE)&-AES_BLOCK_SIZE)
 784                                 - len);
 785                         }
 786                 else
 787                         {
 788                         if (arg>13) arg = 13;
 789                         memcpy(key->aux.tls_aad,ptr,arg);
 790                         key->payload_length = arg;
 791
 792                         return SHA256_DIGEST_LENGTH;
 793                         }
 794                 }
 795 #if !defined(OPENSSL_NO_MULTIBLOCK) && EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK
 796         case EVP_CTRL_TLS1_1_MULTIBLOCK_MAX_BUFSIZE:
 797                 return (int)(5+16+((arg+32+16)&-16));
 798         case EVP_CTRL_TLS1_1_MULTIBLOCK_AAD:
 799                 {
 800                 EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *param =
 801                         (EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *)ptr;
 802                 unsigned int n4x=1, x4;
 803                 unsigned int frag, last, packlen, inp_len;
 804
 805                 if (arg < 0)
 806                         return -1;
 807
 808                 if (u_arg < sizeof(EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM)) return -1;
 809
 810                 inp_len = param->inp[11]<<8|param->inp[12];
 811
 812                 if (ctx->encrypt)
 813                         {
 814                         if ((param->inp[9]<<8|param->inp[10]) < TLS1_1_VERSION)
 815                                 return -1;
 816
 817                         if (inp_len)
 818                                 {
 819                                 if (inp_len<4096) return 0;     /* too short */
 820
 821                                 if (inp_len>=8192 && OPENSSL_ia32cap_P[2]&(1<<5))
 822                                         n4x=2;  /* AVX2 */
 823                                 }
 824                         else if ((n4x=param->interleave/4) && n4x<=2)
 825                                 inp_len = param->len;
 826                         else
 827                                 return -1;
 828
 829                         key->md = key->head;
 830                         SHA256_Update(&key->md,param->inp,13);
 831
 832                         x4 = 4*n4x; n4x += 1;
 833
 834                         frag = inp_len>>n4x;
 835                         last = inp_len+frag-(frag<<n4x);
 836                         if (last>frag && ((last+13+9)%64<(x4-1))) {
 837                                 frag++;
 838                                 last -= x4-1;
 839                         }
 840
 841                         packlen = 5+16+((frag+32+16)&-16);
 842                         packlen = (packlen<<n4x)-packlen;
 843                         packlen += 5+16+((last+32+16)&-16);
 844
 845                         param->interleave = x4;
 846
 847                         return (int)packlen;
 848                         }
 849                 else
 850                         return -1;      /* not yet */
 851                 }
 852         case EVP_CTRL_TLS1_1_MULTIBLOCK_ENCRYPT:
 853                 {
 854                 EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *param =
 855                         (EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *)ptr;
 856
 857                 return (int)tls1_1_multi_block_encrypt(key,param->out,param->inp,
 858                                                 param->len,param->interleave/4);
 859                 }
 860         case EVP_CTRL_TLS1_1_MULTIBLOCK_DECRYPT:
 861 #endif
 862         default:
 863                 return -1;
 864                 }
 865         }
 866
 867 static EVP_CIPHER aesni_128_cbc_hmac_sha256_cipher =
 868         {
 869 #ifdef NID_aes_128_cbc_hmac_sha256
 870         NID_aes_128_cbc_hmac_sha256,
 871 #else
 872         NID_undef,
 873 #endif
 874         16,16,16,
 875         EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|
 876         EVP_CIPH_FLAG_AEAD_CIPHER|EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK,
 877         aesni_cbc_hmac_sha256_init_key,
 878         aesni_cbc_hmac_sha256_cipher,
 879         NULL,
 880         sizeof(EVP_AES_HMAC_SHA256),
 881         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_set_asn1_iv,
 882         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_get_asn1_iv,
 883         aesni_cbc_hmac_sha256_ctrl,
 884         NULL
 885         };
 886
 887 static EVP_CIPHER aesni_256_cbc_hmac_sha256_cipher =
 888         {
 889 #ifdef NID_aes_256_cbc_hmac_sha256
 890         NID_aes_256_cbc_hmac_sha256,
 891 #else
 892         NID_undef,
 893 #endif
 894         16,32,16,
 895         EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|
 896         EVP_CIPH_FLAG_AEAD_CIPHER|EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK,
 897         aesni_cbc_hmac_sha256_init_key,
 898         aesni_cbc_hmac_sha256_cipher,
 899         NULL,
 900         sizeof(EVP_AES_HMAC_SHA256),
 901         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_set_asn1_iv,
 902         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_get_asn1_iv,
 903         aesni_cbc_hmac_sha256_ctrl,
 904         NULL
 905         };
 906
 907 const EVP_CIPHER *EVP_aes_128_cbc_hmac_sha256(void)
 908         {
 909         return((OPENSSL_ia32cap_P[1]&AESNI_CAPABLE) &&
 910                 aesni_cbc_sha256_enc(NULL,NULL,0,NULL,NULL,NULL,NULL) ?
 911                 &aesni_128_cbc_hmac_sha256_cipher:NULL);
 912         }
 913
 914 const EVP_CIPHER *EVP_aes_256_cbc_hmac_sha256(void)
 915         {
 916         return((OPENSSL_ia32cap_P[1]&AESNI_CAPABLE) &&
 917                 aesni_cbc_sha256_enc(NULL,NULL,0,NULL,NULL,NULL,NULL)?
 918                 &aesni_256_cbc_hmac_sha256_cipher:NULL);
 919         }
 920 #else
 921 const EVP_CIPHER *EVP_aes_128_cbc_hmac_sha256(void)
 922         {
 923         return NULL;
 924         }
 925 const EVP_CIPHER *EVP_aes_256_cbc_hmac_sha256(void)
 926         {
 927         return NULL;
 928         }
 929 #endif
 930 #endif