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;
 205         size_t          ret = 0;
 206         u8              *IVs;
 207 #if defined(BSWAP8)
 208         u64             seqnum;
 209 #endif
 210
 211         ctx = (SHA256_MB_CTX *)(storage+32-((size_t)storage%32));       /* align */
 212
 213         frag = (unsigned int)inp_len>>(1+n4x);
 214         last = (unsigned int)inp_len+frag-(frag<<(1+n4x));
 215         if (last>frag && ((last+13+9)%64)<(x4-1)) {
 216                 frag++;
 217                 last -= x4-1;
 218         }
 219
 220         hash_d[0].ptr = inp;
 221         for (i=1;i<x4;i++)      hash_d[i].ptr = hash_d[i-1].ptr+frag;
 222
 223 #if defined(BSWAP8)
 224         memcpy(blocks[0].c,key->md.data,8);
 225         seqnum = BSWAP8(blocks[0].q[0]);
 226 #endif
 227         for (i=0;i<x4;i++) {
 228                 unsigned int len = (i==(x4-1)?last:frag);
 229
 230                 ctx->A[i] = key->md.h[0];
 231                 ctx->B[i] = key->md.h[1];
 232                 ctx->C[i] = key->md.h[2];
 233                 ctx->D[i] = key->md.h[3];
 234                 ctx->E[i] = key->md.h[4];
 235                 ctx->F[i] = key->md.h[5];
 236                 ctx->G[i] = key->md.h[6];
 237                 ctx->H[i] = key->md.h[7];
 238
 239                 /* fix seqnum */
 240 #if defined(BSWAP8)
 241                 blocks[i].q[0] = BSWAP8(seqnum+i);
 242 #else
 243                 blocks[i].c[7] += ((u8*)key->md.data)[7]+i;
 244                 if (blocks[i].c[7] < i) {
 245                         int j;
 246
 247                         for (j=6;j>=0;j--) {
 248                                 if (blocks[i].c[j]=((u8*)key->md.data)[j]+1) break;
 249                         }
 250                 }
 251 #endif
 252                 blocks[i].c[8] = ((u8*)key->md.data)[8];
 253                 blocks[i].c[9] = ((u8*)key->md.data)[9];
 254                 blocks[i].c[10] = ((u8*)key->md.data)[10];
 255                 /* fix length */
 256                 blocks[i].c[11] = (u8)(len>>8);
 257                 blocks[i].c[12] = (u8)(len);
 258
 259                 memcpy(blocks[i].c+13,hash_d[i].ptr,64-13);
 260                 hash_d[i].ptr += 64-13;
 261                 hash_d[i].blocks = (len-(64-13))/64;
 262
 263                 edges[i].ptr = blocks[i].c;
 264                 edges[i].blocks = 1;
 265         }
 266
 267         /* hash 13-byte headers and first 64-13 bytes of inputs */
 268         sha256_multi_block(ctx,edges,n4x);
 269         /* hash bulk inputs */
 270         sha256_multi_block(ctx,hash_d,n4x);
 271
 272         memset(blocks,0,sizeof(blocks));
 273         for (i=0;i<x4;i++) {
 274                 unsigned int            len = (i==(x4-1)?last:frag),
 275                                         off = hash_d[i].blocks*64;
 276                 const unsigned char    *ptr = hash_d[i].ptr+off;
 277
 278                 off = len-(64-13)-off;  /* remainder actually */
 279                 memcpy(blocks[i].c,ptr,off);
 280                 blocks[i].c[off]=0x80;
 281                 len += 64+13;           /* 64 is HMAC header */
 282                 len *= 8;               /* convert to bits */
 283                 if (off<(64-8)) {
 284                         blocks[i].d[15] = BSWAP4(len);
 285                         edges[i].blocks = 1;
 286                 } else {
 287                         blocks[i].d[31] = BSWAP4(len);
 288                         edges[i].blocks = 2;
 289                 }
 290                 edges[i].ptr = blocks[i].c;
 291         }
 292
 293         /* hash input tails and finalize */
 294         sha256_multi_block(ctx,edges,n4x);
 295
 296         memset(blocks,0,sizeof(blocks));
 297         for (i=0;i<x4;i++) {
 298                 blocks[i].d[0] = BSWAP4(ctx->A[i]);     ctx->A[i] = key->tail.h[0];
 299                 blocks[i].d[1] = BSWAP4(ctx->B[i]);     ctx->B[i] = key->tail.h[1];
 300                 blocks[i].d[2] = BSWAP4(ctx->C[i]);     ctx->C[i] = key->tail.h[2];
 301                 blocks[i].d[3] = BSWAP4(ctx->D[i]);     ctx->D[i] = key->tail.h[3];
 302                 blocks[i].d[4] = BSWAP4(ctx->E[i]);     ctx->E[i] = key->tail.h[4];
 303                 blocks[i].d[5] = BSWAP4(ctx->F[i]);     ctx->F[i] = key->tail.h[5];
 304                 blocks[i].d[6] = BSWAP4(ctx->G[i]);     ctx->G[i] = key->tail.h[6];
 305                 blocks[i].d[7] = BSWAP4(ctx->H[i]);     ctx->H[i] = key->tail.h[7];
 306                 blocks[i].c[32] = 0x80;
 307                 blocks[i].d[15] = BSWAP4((64+32)*8);
 308                 edges[i].ptr = blocks[i].c;
 309                 edges[i].blocks = 1;
 310         }
 311
 312         /* finalize MACs */
 313         sha256_multi_block(ctx,edges,n4x);
 314
 315         packlen = 5+16+((frag+32+16)&-16);
 316
 317         out += (packlen<<(1+n4x))-packlen;
 318         inp += (frag<<(1+n4x))-frag;
 319
 320         RAND_bytes((IVs=blocks[0].c),16*x4);    /* ask for IVs in bulk */
 321
 322         for (i=x4-1;;i--) {
 323                 unsigned int len = (i==(x4-1)?last:frag), pad, j;
 324                 unsigned char *out0 = out;
 325
 326                 out += 5+16;            /* place for header and explicit IV */
 327                 ciph_d[i].inp = out;
 328                 ciph_d[i].out = out;
 329
 330                 memmove(out,inp,len);
 331                 out += len;
 332
 333                 /* write MAC */
 334                 ((u32 *)out)[0] = BSWAP4(ctx->A[i]);
 335                 ((u32 *)out)[1] = BSWAP4(ctx->B[i]);
 336                 ((u32 *)out)[2] = BSWAP4(ctx->C[i]);
 337                 ((u32 *)out)[3] = BSWAP4(ctx->D[i]);
 338                 ((u32 *)out)[4] = BSWAP4(ctx->E[i]);
 339                 ((u32 *)out)[5] = BSWAP4(ctx->F[i]);
 340                 ((u32 *)out)[6] = BSWAP4(ctx->G[i]);
 341                 ((u32 *)out)[7] = BSWAP4(ctx->H[i]);
 342                 out += 32;
 343                 len += 32;
 344
 345                 /* pad */
 346                 pad = 15-len%16;
 347                 for (j=0;j<=pad;j++) *(out++) = pad;
 348                 len += pad+1;
 349
 350                 ciph_d[i].blocks = len/16;
 351                 len += 16;      /* account for explicit iv */
 352
 353                 /* arrange header */
 354                 out0[0] = ((u8*)key->md.data)[8];
 355                 out0[1] = ((u8*)key->md.data)[9];
 356                 out0[2] = ((u8*)key->md.data)[10];
 357                 out0[3] = (u8)(len>>8);
 358                 out0[4] = (u8)(len);
 359
 360                 /* explicit iv */
 361                 memcpy(ciph_d[i].iv, IVs, 16);
 362                 memcpy(&out0[5],     IVs, 16);
 363
 364                 ret += len+5;
 365
 366                 if (i==0) break;
 367
 368                 out = out0-packlen;
 369                 inp -= frag;
 370                 IVs += 16;
 371         }
 372
 373         aesni_multi_cbc_encrypt(ciph_d,&key->ks,n4x);
 374
 375         OPENSSL_cleanse(blocks,sizeof(blocks));
 376         OPENSSL_cleanse(ctx,sizeof(*ctx));
 377
 378         return ret;
 379 }
 380 #endif
 381
 382 static int aesni_cbc_hmac_sha256_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
 383                       const unsigned char *in, size_t len)
 384         {
 385         EVP_AES_HMAC_SHA256 *key = data(ctx);
 386         unsigned int l;
 387         size_t  plen = key->payload_length,
 388                 iv = 0,         /* explicit IV in TLS 1.1 and later */
 389                 sha_off = 0;
 390 #if defined(STITCHED_CALL)
 391         size_t  aes_off = 0,
 392                 blocks;
 393
 394         sha_off = SHA256_CBLOCK-key->md.num;
 395 #endif
 396
 397         key->payload_length = NO_PAYLOAD_LENGTH;
 398
 399         if (len%AES_BLOCK_SIZE) return 0;
 400
 401         if (ctx->encrypt) {
 402                 if (plen==NO_PAYLOAD_LENGTH)
 403                         plen = len;
 404                 else if (len!=((plen+SHA256_DIGEST_LENGTH+AES_BLOCK_SIZE)&-AES_BLOCK_SIZE))
 405                         return 0;
 406                 else if (key->aux.tls_ver >= TLS1_1_VERSION)
 407                         iv = AES_BLOCK_SIZE;
 408
 409 #if defined(STITCHED_CALL)
 410                 if (OPENSSL_ia32cap_P[1]&(1<<(60-32)) && /* AVX? */
 411                     plen>(sha_off+iv) &&
 412                     (blocks=(plen-(sha_off+iv))/SHA256_CBLOCK)) {
 413                         SHA256_Update(&key->md,in+iv,sha_off);
 414
 415                         (void)aesni_cbc_sha256_enc(in,out,blocks,&key->ks,
 416                                 ctx->iv,&key->md,in+iv+sha_off);
 417                         blocks *= SHA256_CBLOCK;
 418                         aes_off += blocks;
 419                         sha_off += blocks;
 420                         key->md.Nh += blocks>>29;
 421                         key->md.Nl += blocks<<=3;
 422                         if (key->md.Nl<(unsigned int)blocks) key->md.Nh++;
 423                 } else {
 424                         sha_off = 0;
 425                 }
 426 #endif
 427                 sha_off += iv;
 428                 SHA256_Update(&key->md,in+sha_off,plen-sha_off);
 429
 430                 if (plen!=len)  {       /* "TLS" mode of operation */
 431                         if (in!=out)
 432                                 memcpy(out+aes_off,in+aes_off,plen-aes_off);
 433
 434                         /* calculate HMAC and append it to payload */
 435                         SHA256_Final(out+plen,&key->md);
 436                         key->md = key->tail;
 437                         SHA256_Update(&key->md,out+plen,SHA256_DIGEST_LENGTH);
 438                         SHA256_Final(out+plen,&key->md);
 439
 440                         /* pad the payload|hmac */
 441                         plen += SHA256_DIGEST_LENGTH;
 442                         for (l=len-plen-1;plen<len;plen++) out[plen]=l;
 443                         /* encrypt HMAC|padding at once */
 444                         aesni_cbc_encrypt(out+aes_off,out+aes_off,len-aes_off,
 445                                         &key->ks,ctx->iv,1);
 446                 } else {
 447                         aesni_cbc_encrypt(in+aes_off,out+aes_off,len-aes_off,
 448                                         &key->ks,ctx->iv,1);
 449                 }
 450         } else {
 451                 union { unsigned int  u[SHA256_DIGEST_LENGTH/sizeof(unsigned int)];
 452                         unsigned char c[64+SHA256_DIGEST_LENGTH]; } mac, *pmac;
 453
 454                 /* arrange cache line alignment */
 455                 pmac = (void *)(((size_t)mac.c+63)&((size_t)0-64));
 456
 457                 /* decrypt HMAC|padding at once */
 458                 aesni_cbc_encrypt(in,out,len,
 459                                 &key->ks,ctx->iv,0);
 460
 461                 if (plen != NO_PAYLOAD_LENGTH) {        /* "TLS" mode of operation */
 462                         size_t inp_len, mask, j, i;
 463                         unsigned int res, maxpad, pad, bitlen;
 464                         int ret = 1;
 465                         union { unsigned int  u[SHA_LBLOCK];
 466                                 unsigned char c[SHA256_CBLOCK]; }
 467                                 *data = (void *)key->md.data;
 468
 469                         if ((key->aux.tls_aad[plen-4]<<8|key->aux.tls_aad[plen-3])
 470                             >= TLS1_1_VERSION)
 471                                 iv = AES_BLOCK_SIZE;
 472
 473                         if (len<(iv+SHA256_DIGEST_LENGTH+1))
 474                                 return 0;
 475
 476                         /* omit explicit iv */
 477                         out += iv;
 478                         len -= iv;
 479
 480                         /* figure out payload length */
 481                         pad = out[len-1];
 482                         maxpad = len-(SHA256_DIGEST_LENGTH+1);
 483                         maxpad |= (255-maxpad)>>(sizeof(maxpad)*8-8);
 484                         maxpad &= 255;
 485
 486                         inp_len = len - (SHA256_DIGEST_LENGTH+pad+1);
 487                         mask = (0-((inp_len-len)>>(sizeof(inp_len)*8-1)));
 488                         inp_len &= mask;
 489                         ret &= (int)mask;
 490
 491                         key->aux.tls_aad[plen-2] = inp_len>>8;
 492                         key->aux.tls_aad[plen-1] = inp_len;
 493
 494                         /* calculate HMAC */
 495                         key->md = key->head;
 496                         SHA256_Update(&key->md,key->aux.tls_aad,plen);
 497
 498 #if 1
 499                         len -= SHA256_DIGEST_LENGTH;            /* amend mac */
 500                         if (len>=(256+SHA256_CBLOCK)) {
 501                                 j = (len-(256+SHA256_CBLOCK))&(0-SHA256_CBLOCK);
 502                                 j += SHA256_CBLOCK-key->md.num;
 503                                 SHA256_Update(&key->md,out,j);
 504                                 out += j;
 505                                 len -= j;
 506                                 inp_len -= j;
 507                         }
 508
 509                         /* but pretend as if we hashed padded payload */
 510                         bitlen = key->md.Nl+(inp_len<<3);       /* at most 18 bits */
 511 #ifdef BSWAP4
 512                         bitlen = BSWAP4(bitlen);
 513 #else
 514                         mac.c[0] = 0;
 515                         mac.c[1] = (unsigned char)(bitlen>>16);
 516                         mac.c[2] = (unsigned char)(bitlen>>8);
 517                         mac.c[3] = (unsigned char)bitlen;
 518                         bitlen = mac.u[0];
 519 #endif
 520
 521                         pmac->u[0]=0;
 522                         pmac->u[1]=0;
 523                         pmac->u[2]=0;
 524                         pmac->u[3]=0;
 525                         pmac->u[4]=0;
 526                         pmac->u[5]=0;
 527                         pmac->u[6]=0;
 528                         pmac->u[7]=0;
 529
 530                         for (res=key->md.num, j=0;j<len;j++) {
 531                                 size_t c = out[j];
 532                                 mask = (j-inp_len)>>(sizeof(j)*8-8);
 533                                 c &= mask;
 534                                 c |= 0x80&~mask&~((inp_len-j)>>(sizeof(j)*8-8));
 535                                 data->c[res++]=(unsigned char)c;
 536
 537                                 if (res!=SHA256_CBLOCK) continue;
 538
 539                                 /* j is not incremented yet */
 540                                 mask = 0-((inp_len+7-j)>>(sizeof(j)*8-1));
 541                                 data->u[SHA_LBLOCK-1] |= bitlen&mask;
 542                                 sha256_block_data_order(&key->md,data,1);
 543                                 mask &= 0-((j-inp_len-72)>>(sizeof(j)*8-1));
 544                                 pmac->u[0] |= key->md.h[0] & mask;
 545                                 pmac->u[1] |= key->md.h[1] & mask;
 546                                 pmac->u[2] |= key->md.h[2] & mask;
 547                                 pmac->u[3] |= key->md.h[3] & mask;
 548                                 pmac->u[4] |= key->md.h[4] & mask;
 549                                 pmac->u[5] |= key->md.h[5] & mask;
 550                                 pmac->u[6] |= key->md.h[6] & mask;
 551                                 pmac->u[7] |= key->md.h[7] & mask;
 552                                 res=0;
 553                         }
 554
 555                         for(i=res;i<SHA256_CBLOCK;i++,j++) data->c[i]=0;
 556
 557                         if (res>SHA256_CBLOCK-8) {
 558                                 mask = 0-((inp_len+8-j)>>(sizeof(j)*8-1));
 559                                 data->u[SHA_LBLOCK-1] |= bitlen&mask;
 560                                 sha256_block_data_order(&key->md,data,1);
 561                                 mask &= 0-((j-inp_len-73)>>(sizeof(j)*8-1));
 562                                 pmac->u[0] |= key->md.h[0] & mask;
 563                                 pmac->u[1] |= key->md.h[1] & mask;
 564                                 pmac->u[2] |= key->md.h[2] & mask;
 565                                 pmac->u[3] |= key->md.h[3] & mask;
 566                                 pmac->u[4] |= key->md.h[4] & mask;
 567                                 pmac->u[5] |= key->md.h[5] & mask;
 568                                 pmac->u[6] |= key->md.h[6] & mask;
 569                                 pmac->u[7] |= key->md.h[7] & mask;
 570
 571                                 memset(data,0,SHA256_CBLOCK);
 572                                 j+=64;
 573                         }
 574                         data->u[SHA_LBLOCK-1] = bitlen;
 575                         sha256_block_data_order(&key->md,data,1);
 576                         mask = 0-((j-inp_len-73)>>(sizeof(j)*8-1));
 577                         pmac->u[0] |= key->md.h[0] & mask;
 578                         pmac->u[1] |= key->md.h[1] & mask;
 579                         pmac->u[2] |= key->md.h[2] & mask;
 580                         pmac->u[3] |= key->md.h[3] & mask;
 581                         pmac->u[4] |= key->md.h[4] & mask;
 582                         pmac->u[5] |= key->md.h[5] & mask;
 583                         pmac->u[6] |= key->md.h[6] & mask;
 584                         pmac->u[7] |= key->md.h[7] & mask;
 585
 586 #ifdef BSWAP4
 587                         pmac->u[0] = BSWAP4(pmac->u[0]);
 588                         pmac->u[1] = BSWAP4(pmac->u[1]);
 589                         pmac->u[2] = BSWAP4(pmac->u[2]);
 590                         pmac->u[3] = BSWAP4(pmac->u[3]);
 591                         pmac->u[4] = BSWAP4(pmac->u[4]);
 592                         pmac->u[5] = BSWAP4(pmac->u[5]);
 593                         pmac->u[6] = BSWAP4(pmac->u[6]);
 594                         pmac->u[7] = BSWAP4(pmac->u[7]);
 595 #else
 596                         for (i=0;i<8;i++) {
 597                                 res = pmac->u[i];
 598                                 pmac->c[4*i+0]=(unsigned char)(res>>24);
 599                                 pmac->c[4*i+1]=(unsigned char)(res>>16);
 600                                 pmac->c[4*i+2]=(unsigned char)(res>>8);
 601                                 pmac->c[4*i+3]=(unsigned char)res;
 602                         }
 603 #endif
 604                         len += SHA256_DIGEST_LENGTH;
 605 #else
 606                         SHA256_Update(&key->md,out,inp_len);
 607                         res = key->md.num;
 608                         SHA256_Final(pmac->c,&key->md);
 609
 610                         {
 611                         unsigned int inp_blocks, pad_blocks;
 612
 613                         /* but pretend as if we hashed padded payload */
 614                         inp_blocks = 1+((SHA256_CBLOCK-9-res)>>(sizeof(res)*8-1));
 615                         res += (unsigned int)(len-inp_len);
 616                         pad_blocks = res / SHA256_CBLOCK;
 617                         res %= SHA256_CBLOCK;
 618                         pad_blocks += 1+((SHA256_CBLOCK-9-res)>>(sizeof(res)*8-1));
 619                         for (;inp_blocks<pad_blocks;inp_blocks++)
 620                                 sha1_block_data_order(&key->md,data,1);
 621                         }
 622 #endif
 623                         key->md = key->tail;
 624                         SHA256_Update(&key->md,pmac->c,SHA256_DIGEST_LENGTH);
 625                         SHA256_Final(pmac->c,&key->md);
 626
 627                         /* verify HMAC */
 628                         out += inp_len;
 629                         len -= inp_len;
 630 #if 1
 631                         {
 632                         unsigned char *p = out+len-1-maxpad-SHA256_DIGEST_LENGTH;
 633                         size_t off = out-p;
 634                         unsigned int c, cmask;
 635
 636                         maxpad += SHA256_DIGEST_LENGTH;
 637                         for (res=0,i=0,j=0;j<maxpad;j++) {
 638                                 c = p[j];
 639                                 cmask = ((int)(j-off-SHA256_DIGEST_LENGTH))>>(sizeof(int)*8-1);
 640                                 res |= (c^pad)&~cmask;  /* ... and padding */
 641                                 cmask &= ((int)(off-1-j))>>(sizeof(int)*8-1);
 642                                 res |= (c^pmac->c[i])&cmask;
 643                                 i += 1&cmask;
 644                         }
 645                         maxpad -= SHA256_DIGEST_LENGTH;
 646
 647                         res = 0-((0-res)>>(sizeof(res)*8-1));
 648                         ret &= (int)~res;
 649                         }
 650 #else
 651                         for (res=0,i=0;i<SHA256_DIGEST_LENGTH;i++)
 652                                 res |= out[i]^pmac->c[i];
 653                         res = 0-((0-res)>>(sizeof(res)*8-1));
 654                         ret &= (int)~res;
 655
 656                         /* verify padding */
 657                         pad = (pad&~res) | (maxpad&res);
 658                         out = out+len-1-pad;
 659                         for (res=0,i=0;i<pad;i++)
 660                                 res |= out[i]^pad;
 661
 662                         res = (0-res)>>(sizeof(res)*8-1);
 663                         ret &= (int)~res;
 664 #endif
 665                         return ret;
 666                 } else {
 667                         SHA256_Update(&key->md,out,len);
 668                 }
 669         }
 670
 671         return 1;
 672         }
 673
 674 static int aesni_cbc_hmac_sha256_ctrl(EVP_CIPHER_CTX *ctx, int type, int arg, void *ptr)
 675         {
 676         EVP_AES_HMAC_SHA256 *key = data(ctx);
 677
 678         switch (type)
 679                 {
 680         case EVP_CTRL_AEAD_SET_MAC_KEY:
 681                 {
 682                 unsigned int  i;
 683                 unsigned char hmac_key[64];
 684
 685                 memset (hmac_key,0,sizeof(hmac_key));
 686
 687                 if (arg > (int)sizeof(hmac_key)) {
 688                         SHA256_Init(&key->head);
 689                         SHA256_Update(&key->head,ptr,arg);
 690                         SHA256_Final(hmac_key,&key->head);
 691                 } else {
 692                         memcpy(hmac_key,ptr,arg);
 693                 }
 694
 695                 for (i=0;i<sizeof(hmac_key);i++)
 696                         hmac_key[i] ^= 0x36;            /* ipad */
 697                 SHA256_Init(&key->head);
 698                 SHA256_Update(&key->head,hmac_key,sizeof(hmac_key));
 699
 700                 for (i=0;i<sizeof(hmac_key);i++)
 701                         hmac_key[i] ^= 0x36^0x5c;       /* opad */
 702                 SHA256_Init(&key->tail);
 703                 SHA256_Update(&key->tail,hmac_key,sizeof(hmac_key));
 704
 705                 OPENSSL_cleanse(hmac_key,sizeof(hmac_key));
 706
 707                 return 1;
 708                 }
 709         case EVP_CTRL_AEAD_TLS1_AAD:
 710                 {
 711                 unsigned char *p=ptr;
 712                 unsigned int   len=p[arg-2]<<8|p[arg-1];
 713
 714                 if (ctx->encrypt)
 715                         {
 716                         key->payload_length = len;
 717                         if ((key->aux.tls_ver=p[arg-4]<<8|p[arg-3]) >= TLS1_1_VERSION) {
 718                                 len -= AES_BLOCK_SIZE;
 719                                 p[arg-2] = len>>8;
 720                                 p[arg-1] = len;
 721                         }
 722                         key->md = key->head;
 723                         SHA256_Update(&key->md,p,arg);
 724
 725                         return (int)(((len+SHA256_DIGEST_LENGTH+AES_BLOCK_SIZE)&-AES_BLOCK_SIZE)
 726                                 - len);
 727                         }
 728                 else
 729                         {
 730                         if (arg>13) arg = 13;
 731                         memcpy(key->aux.tls_aad,ptr,arg);
 732                         key->payload_length = arg;
 733
 734                         return SHA256_DIGEST_LENGTH;
 735                         }
 736                 }
 737 #if !defined(OPENSSL_NO_MULTIBLOCK) && EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK
 738         case EVP_CTRL_TLS1_1_MULTIBLOCK_MAX_BUFSIZE:
 739                 return (int)(5+16+((arg+32+16)&-16));
 740         case EVP_CTRL_TLS1_1_MULTIBLOCK_AAD:
 741                 {
 742                 EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *param =
 743                         (EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *)ptr;
 744                 unsigned int n4x=1, x4;
 745                 unsigned int frag, last, packlen, inp_len;
 746
 747                 if (arg<sizeof(EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM)) return -1;
 748
 749                 inp_len = param->inp[11]<<8|param->inp[12];
 750
 751                 if (ctx->encrypt)
 752                         {
 753                         if ((param->inp[9]<<8|param->inp[10]) < TLS1_1_VERSION)
 754                                 return -1;
 755
 756                         if (inp_len)
 757                                 {
 758                                 if (inp_len<4096) return 0;     /* too short */
 759
 760                                 if (inp_len>=8192 && OPENSSL_ia32cap_P[2]&(1<<5))
 761                                         n4x=2;  /* AVX2 */
 762                                 }
 763                         else if ((n4x=param->interleave/4) && n4x<=2)
 764                                 inp_len = param->len;
 765                         else
 766                                 return -1;
 767
 768                         key->md = key->head;
 769                         SHA256_Update(&key->md,param->inp,13);
 770
 771                         x4 = 4*n4x; n4x += 1;
 772
 773                         frag = inp_len>>n4x;
 774                         last = inp_len+frag-(frag<<n4x);
 775                         if (last>frag && ((last+13+9)%64<(x4-1))) {
 776                                 frag++;
 777                                 last -= x4-1;
 778                         }
 779
 780                         packlen = 5+16+((frag+32+16)&-16);
 781                         packlen = (packlen<<n4x)-packlen;
 782                         packlen += 5+16+((last+32+16)&-16);
 783
 784                         param->interleave = x4;
 785
 786                         return (int)packlen;
 787                         }
 788                 else
 789                         return -1;      /* not yet */
 790                 }
 791         case EVP_CTRL_TLS1_1_MULTIBLOCK_ENCRYPT:
 792                 {
 793                 EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *param =
 794                         (EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *)ptr;
 795
 796                 return (int)tls1_1_multi_block_encrypt(key,param->out,param->inp,
 797                                                 param->len,param->interleave/4);
 798                 }
 799         case EVP_CTRL_TLS1_1_MULTIBLOCK_DECRYPT:
 800 #endif
 801         default:
 802                 return -1;
 803                 }
 804         }
 805
 806 static EVP_CIPHER aesni_128_cbc_hmac_sha256_cipher =
 807         {
 808 #ifdef NID_aes_128_cbc_hmac_sha256
 809         NID_aes_128_cbc_hmac_sha256,
 810 #else
 811         NID_undef,
 812 #endif
 813         16,16,16,
 814         EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|
 815         EVP_CIPH_FLAG_AEAD_CIPHER|EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK,
 816         aesni_cbc_hmac_sha256_init_key,
 817         aesni_cbc_hmac_sha256_cipher,
 818         NULL,
 819         sizeof(EVP_AES_HMAC_SHA256),
 820         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_set_asn1_iv,
 821         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_get_asn1_iv,
 822         aesni_cbc_hmac_sha256_ctrl,
 823         NULL
 824         };
 825
 826 static EVP_CIPHER aesni_256_cbc_hmac_sha256_cipher =
 827         {
 828 #ifdef NID_aes_256_cbc_hmac_sha256
 829         NID_aes_256_cbc_hmac_sha256,
 830 #else
 831         NID_undef,
 832 #endif
 833         16,32,16,
 834         EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|
 835         EVP_CIPH_FLAG_AEAD_CIPHER|EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK,
 836         aesni_cbc_hmac_sha256_init_key,
 837         aesni_cbc_hmac_sha256_cipher,
 838         NULL,
 839         sizeof(EVP_AES_HMAC_SHA256),
 840         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_set_asn1_iv,
 841         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_get_asn1_iv,
 842         aesni_cbc_hmac_sha256_ctrl,
 843         NULL
 844         };
 845
 846 const EVP_CIPHER *EVP_aes_128_cbc_hmac_sha256(void)
 847         {
 848         return((OPENSSL_ia32cap_P[1]&AESNI_CAPABLE) &&
 849                 aesni_cbc_sha256_enc(NULL,NULL,0,NULL,NULL,NULL,NULL) ?
 850                 &aesni_128_cbc_hmac_sha256_cipher:NULL);
 851         }
 852
 853 const EVP_CIPHER *EVP_aes_256_cbc_hmac_sha256(void)
 854         {
 855         return((OPENSSL_ia32cap_P[1]&AESNI_CAPABLE) &&
 856                 aesni_cbc_sha256_enc(NULL,NULL,0,NULL,NULL,NULL,NULL)?
 857                 &aesni_256_cbc_hmac_sha256_cipher:NULL);
 858         }
 859 #else
 860 const EVP_CIPHER *EVP_aes_128_cbc_hmac_sha256(void)
 861         {
 862         return NULL;
 863         }
 864 const EVP_CIPHER *EVP_aes_256_cbc_hmac_sha256(void)
 865         {
 866         return NULL;
 867         }
 868 #endif
 869 #endif