crypto/evp/e_aes_cbc_hmac_sha1.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_SHA1)
  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     SHA_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_SHA1;
  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 void aesni_cbc_sha1_enc (const void *inp, void *out, size_t blocks,
 113                 const AES_KEY *key, unsigned char iv[16],
 114                 SHA_CTX *ctx,const void *in0);
 115
 116 void aesni256_cbc_sha1_dec (const void *inp, void *out, size_t blocks,
 117                 const AES_KEY *key, unsigned char iv[16],
 118                 SHA_CTX *ctx,const void *in0);
 119
 120 #define data(ctx) ((EVP_AES_HMAC_SHA1 *)(ctx)->cipher_data)
 121
 122 static int aesni_cbc_hmac_sha1_init_key(EVP_CIPHER_CTX *ctx,
 123                         const unsigned char *inkey,
 124                         const unsigned char *iv, int enc)
 125         {
 126         EVP_AES_HMAC_SHA1 *key = data(ctx);
 127         int ret;
 128
 129         if (enc)
 130                 ret=aesni_set_encrypt_key(inkey,ctx->key_len*8,&key->ks);
 131         else
 132                 ret=aesni_set_decrypt_key(inkey,ctx->key_len*8,&key->ks);
 133
 134         SHA1_Init(&key->head);  /* handy when benchmarking */
 135         key->tail = key->head;
 136         key->md   = key->head;
 137
 138         key->payload_length = NO_PAYLOAD_LENGTH;
 139
 140         return ret<0?0:1;
 141         }
 142
 143 #define STITCHED_CALL
 144 #undef  STITCHED_DECRYPT_CALL
 145
 146 #if !defined(STITCHED_CALL)
 147 #define aes_off 0
 148 #endif
 149
 150 void sha1_block_data_order (void *c,const void *p,size_t len);
 151
 152 static void sha1_update(SHA_CTX *c,const void *data,size_t len)
 153 {       const unsigned char *ptr = data;
 154         size_t res;
 155
 156         if ((res = c->num)) {
 157                 res = SHA_CBLOCK-res;
 158                 if (len<res) res=len;
 159                 SHA1_Update (c,ptr,res);
 160                 ptr += res;
 161                 len -= res;
 162         }
 163
 164         res = len % SHA_CBLOCK;
 165         len -= res;
 166
 167         if (len) {
 168                 sha1_block_data_order(c,ptr,len/SHA_CBLOCK);
 169
 170                 ptr += len;
 171                 c->Nh += len>>29;
 172                 c->Nl += len<<=3;
 173                 if (c->Nl<(unsigned int)len) c->Nh++;
 174         }
 175
 176         if (res)
 177                 SHA1_Update(c,ptr,res);
 178 }
 179
 180 #ifdef SHA1_Update
 181 #undef SHA1_Update
 182 #endif
 183 #define SHA1_Update sha1_update
 184
 185 #if !defined(OPENSSL_NO_MULTIBLOCK) && EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK
 186
 187 typedef struct { unsigned int A[8],B[8],C[8],D[8],E[8]; } SHA1_MB_CTX;
 188 typedef struct { const unsigned char *ptr; int blocks;  } HASH_DESC;
 189
 190 void sha1_multi_block(SHA1_MB_CTX *,const HASH_DESC *,int);
 191
 192 typedef struct { const unsigned char *inp; unsigned char *out;
 193                  int blocks; u64 iv[2]; } CIPH_DESC;
 194
 195 void aesni_multi_cbc_encrypt(CIPH_DESC *,void *,int);
 196
 197 static size_t tls1_1_multi_block_encrypt(EVP_AES_HMAC_SHA1 *key,
 198         unsigned char *out, const unsigned char *inp, size_t inp_len,
 199         int n4x)        /* n4x is 1 or 2 */
 200 {
 201         HASH_DESC       hash_d[8], edges[8];
 202         CIPH_DESC       ciph_d[8];
 203         unsigned char   storage[sizeof(SHA1_MB_CTX)+32];
 204         union { u64     q[16];
 205                 u32     d[32];
 206                 u8      c[128]; } blocks[8];
 207         SHA1_MB_CTX     *ctx;
 208         unsigned int    frag, last, packlen, i, x4=4*n4x, minblocks, processed=0;
 209         size_t          ret = 0;
 210         u8              *IVs;
 211 #if defined(BSWAP8)
 212         u64             seqnum;
 213 #endif
 214
 215         /* ask for IVs in bulk */
 216         if (RAND_bytes((IVs=blocks[0].c),16*x4)<=0)
 217                 return 0;
 218
 219         ctx = (SHA1_MB_CTX *)(storage+32-((size_t)storage%32)); /* align */
 220
 221         frag = (unsigned int)inp_len>>(1+n4x);
 222         last = (unsigned int)inp_len+frag-(frag<<(1+n4x));
 223         if (last>frag && ((last+13+9)%64)<(x4-1)) {
 224                 frag++;
 225                 last -= x4-1;
 226         }
 227
 228         packlen = 5+16+((frag+20+16)&-16);
 229
 230         /* populate descriptors with pointers and IVs */
 231         hash_d[0].ptr = inp;
 232         ciph_d[0].inp = inp;
 233         /* 5+16 is place for header and explicit IV */
 234         ciph_d[0].out = out+5+16;
 235         memcpy(ciph_d[0].out-16,IVs,16);
 236         memcpy(ciph_d[0].iv,IVs,16);    IVs += 16;
 237
 238         for (i=1;i<x4;i++) {
 239                 ciph_d[i].inp = hash_d[i].ptr = hash_d[i-1].ptr+frag;
 240                 ciph_d[i].out = ciph_d[i-1].out+packlen;
 241                 memcpy(ciph_d[i].out-16,IVs,16);
 242                 memcpy(ciph_d[i].iv,IVs,16);    IVs+=16;
 243         }
 244
 245 #if defined(BSWAP8)
 246         memcpy(blocks[0].c,key->md.data,8);
 247         seqnum = BSWAP8(blocks[0].q[0]);
 248 #endif
 249         for (i=0;i<x4;i++) {
 250                 unsigned int len = (i==(x4-1)?last:frag);
 251 #if !defined(BSWAP8)
 252                 unsigned int carry, j;
 253 #endif
 254
 255                 ctx->A[i] = key->md.h0;
 256                 ctx->B[i] = key->md.h1;
 257                 ctx->C[i] = key->md.h2;
 258                 ctx->D[i] = key->md.h3;
 259                 ctx->E[i] = key->md.h4;
 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         sha1_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                         sha1_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         sha1_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         sha1_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.h0;
 359                 blocks[i].d[1] = BSWAP4(ctx->B[i]);     ctx->B[i] = key->tail.h1;
 360                 blocks[i].d[2] = BSWAP4(ctx->C[i]);     ctx->C[i] = key->tail.h2;
 361                 blocks[i].d[3] = BSWAP4(ctx->D[i]);     ctx->D[i] = key->tail.h3;
 362                 blocks[i].d[4] = BSWAP4(ctx->E[i]);     ctx->E[i] = key->tail.h4;
 363                 blocks[i].c[20] = 0x80;
 364                 blocks[i].d[15] = BSWAP4((64+20)*8);
 365 #else
 366                 PUTU32(blocks[i].c+0,ctx->A[i]);        ctx->A[i] = key->tail.h0;
 367                 PUTU32(blocks[i].c+4,ctx->B[i]);        ctx->B[i] = key->tail.h1;
 368                 PUTU32(blocks[i].c+8,ctx->C[i]);        ctx->C[i] = key->tail.h2;
 369                 PUTU32(blocks[i].c+12,ctx->D[i]);       ctx->D[i] = key->tail.h3;
 370                 PUTU32(blocks[i].c+16,ctx->E[i]);       ctx->E[i] = key->tail.h4;
 371                 blocks[i].c[20] = 0x80;
 372                 PUTU32(blocks[i].c+60,(64+20)*8);
 373 #endif
 374                 edges[i].ptr = blocks[i].c;
 375                 edges[i].blocks = 1;
 376         }
 377
 378         /* finalize MACs */
 379         sha1_multi_block(ctx,edges,n4x);
 380
 381         for (i=0;i<x4;i++) {
 382                 unsigned int len = (i==(x4-1)?last:frag), pad, j;
 383                 unsigned char *out0 = out;
 384
 385                 memcpy(ciph_d[i].out,ciph_d[i].inp,len-processed);
 386                 ciph_d[i].inp = ciph_d[i].out;
 387
 388                 out += 5+16+len;
 389
 390                 /* write MAC */
 391                 PUTU32(out+0,ctx->A[i]);
 392                 PUTU32(out+4,ctx->B[i]);
 393                 PUTU32(out+8,ctx->C[i]);
 394                 PUTU32(out+12,ctx->D[i]);
 395                 PUTU32(out+16,ctx->E[i]);
 396                 out += 20;
 397                 len += 20;
 398
 399                 /* pad */
 400                 pad = 15-len%16;
 401                 for (j=0;j<=pad;j++) *(out++) = pad;
 402                 len += pad+1;
 403
 404                 ciph_d[i].blocks = (len-processed)/16;
 405                 len += 16;      /* account for explicit iv */
 406
 407                 /* arrange header */
 408                 out0[0] = ((u8*)key->md.data)[8];
 409                 out0[1] = ((u8*)key->md.data)[9];
 410                 out0[2] = ((u8*)key->md.data)[10];
 411                 out0[3] = (u8)(len>>8);
 412                 out0[4] = (u8)(len);
 413
 414                 ret += len+5;
 415                 inp += frag;
 416         }
 417
 418         aesni_multi_cbc_encrypt(ciph_d,&key->ks,n4x);
 419
 420         OPENSSL_cleanse(blocks,sizeof(blocks));
 421         OPENSSL_cleanse(ctx,sizeof(*ctx));
 422
 423         return ret;
 424 }
 425 #endif
 426
 427 static int aesni_cbc_hmac_sha1_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
 428                       const unsigned char *in, size_t len)
 429         {
 430         EVP_AES_HMAC_SHA1 *key = data(ctx);
 431         unsigned int l;
 432         size_t  plen = key->payload_length,
 433                 iv = 0,         /* explicit IV in TLS 1.1 and later */
 434                 sha_off = 0;
 435 #if defined(STITCHED_CALL)
 436         size_t  aes_off = 0,
 437                 blocks;
 438
 439         sha_off = SHA_CBLOCK-key->md.num;
 440 #endif
 441
 442         key->payload_length = NO_PAYLOAD_LENGTH;
 443
 444         if (len%AES_BLOCK_SIZE) return 0;
 445
 446         if (ctx->encrypt) {
 447                 if (plen==NO_PAYLOAD_LENGTH)
 448                         plen = len;
 449                 else if (len!=((plen+SHA_DIGEST_LENGTH+AES_BLOCK_SIZE)&-AES_BLOCK_SIZE))
 450                         return 0;
 451                 else if (key->aux.tls_ver >= TLS1_1_VERSION)
 452                         iv = AES_BLOCK_SIZE;
 453
 454 #if defined(STITCHED_CALL)
 455                 if (plen>(sha_off+iv) && (blocks=(plen-(sha_off+iv))/SHA_CBLOCK)) {
 456                         SHA1_Update(&key->md,in+iv,sha_off);
 457
 458                         aesni_cbc_sha1_enc(in,out,blocks,&key->ks,
 459                                 ctx->iv,&key->md,in+iv+sha_off);
 460                         blocks *= SHA_CBLOCK;
 461                         aes_off += blocks;
 462                         sha_off += blocks;
 463                         key->md.Nh += blocks>>29;
 464                         key->md.Nl += blocks<<=3;
 465                         if (key->md.Nl<(unsigned int)blocks) key->md.Nh++;
 466                 } else {
 467                         sha_off = 0;
 468                 }
 469 #endif
 470                 sha_off += iv;
 471                 SHA1_Update(&key->md,in+sha_off,plen-sha_off);
 472
 473                 if (plen!=len)  {       /* "TLS" mode of operation */
 474                         if (in!=out)
 475                                 memcpy(out+aes_off,in+aes_off,plen-aes_off);
 476
 477                         /* calculate HMAC and append it to payload */
 478                         SHA1_Final(out+plen,&key->md);
 479                         key->md = key->tail;
 480                         SHA1_Update(&key->md,out+plen,SHA_DIGEST_LENGTH);
 481                         SHA1_Final(out+plen,&key->md);
 482
 483                         /* pad the payload|hmac */
 484                         plen += SHA_DIGEST_LENGTH;
 485                         for (l=len-plen-1;plen<len;plen++) out[plen]=l;
 486                         /* encrypt HMAC|padding at once */
 487                         aesni_cbc_encrypt(out+aes_off,out+aes_off,len-aes_off,
 488                                         &key->ks,ctx->iv,1);
 489                 } else {
 490                         aesni_cbc_encrypt(in+aes_off,out+aes_off,len-aes_off,
 491                                         &key->ks,ctx->iv,1);
 492                 }
 493         } else {
 494                 union { unsigned int  u[SHA_DIGEST_LENGTH/sizeof(unsigned int)];
 495                         unsigned char c[32+SHA_DIGEST_LENGTH]; } mac, *pmac;
 496
 497                 /* arrange cache line alignment */
 498                 pmac = (void *)(((size_t)mac.c+31)&((size_t)0-32));
 499
 500                 if (plen != NO_PAYLOAD_LENGTH) {        /* "TLS" mode of operation */
 501                         size_t inp_len, mask, j, i;
 502                         unsigned int res, maxpad, pad, bitlen;
 503                         int ret = 1;
 504                         union { unsigned int  u[SHA_LBLOCK];
 505                                 unsigned char c[SHA_CBLOCK]; }
 506                                 *data = (void *)key->md.data;
 507 #if defined(STITCHED_DECRYPT_CALL)
 508                         unsigned char tail_iv[AES_BLOCK_SIZE];
 509                         int stitch=0;
 510 #endif
 511
 512                         if ((key->aux.tls_aad[plen-4]<<8|key->aux.tls_aad[plen-3])
 513                             >= TLS1_1_VERSION) {
 514                                 if (len<(AES_BLOCK_SIZE+SHA_DIGEST_LENGTH+1))
 515                                         return 0;
 516
 517                                 /* omit explicit iv */
 518                                 memcpy(ctx->iv,in,AES_BLOCK_SIZE);
 519                                 in  += AES_BLOCK_SIZE;
 520                                 out += AES_BLOCK_SIZE;
 521                                 len -= AES_BLOCK_SIZE;
 522                         }
 523                         else if (len<(SHA_DIGEST_LENGTH+1))
 524                                 return 0;
 525
 526 #if defined(STITCHED_DECRYPT_CALL)
 527                         if (len>=1024 && ctx->key_len==32) {
 528                                 /* decrypt last block */
 529                                 memcpy(tail_iv,in+len-2*AES_BLOCK_SIZE,AES_BLOCK_SIZE);
 530                                 aesni_cbc_encrypt(in+len-AES_BLOCK_SIZE,
 531                                                 out+len-AES_BLOCK_SIZE,AES_BLOCK_SIZE,
 532                                                 &key->ks,tail_iv,0);
 533                                 stitch=1;
 534                         } else
 535 #endif
 536                         /* decrypt HMAC|padding at once */
 537                         aesni_cbc_encrypt(in,out,len,
 538                                         &key->ks,ctx->iv,0);
 539
 540                         /* figure out payload length */
 541                         pad = out[len-1];
 542                         maxpad = len-(SHA_DIGEST_LENGTH+1);
 543                         maxpad |= (255-maxpad)>>(sizeof(maxpad)*8-8);
 544                         maxpad &= 255;
 545
 546                         inp_len = len - (SHA_DIGEST_LENGTH+pad+1);
 547                         mask = (0-((inp_len-len)>>(sizeof(inp_len)*8-1)));
 548                         inp_len &= mask;
 549                         ret &= (int)mask;
 550
 551                         key->aux.tls_aad[plen-2] = inp_len>>8;
 552                         key->aux.tls_aad[plen-1] = inp_len;
 553
 554                         /* calculate HMAC */
 555                         key->md = key->head;
 556                         SHA1_Update(&key->md,key->aux.tls_aad,plen);
 557
 558 #if defined(STITCHED_DECRYPT_CALL)
 559                         if (stitch) {
 560                                 blocks = (len-(256+32+SHA_CBLOCK))/SHA_CBLOCK;
 561                                 aes_off = len-AES_BLOCK_SIZE-blocks*SHA_CBLOCK;
 562                                 sha_off = SHA_CBLOCK-plen;
 563
 564                                 aesni_cbc_encrypt(in,out,aes_off,
 565                                         &key->ks,ctx->iv,0);
 566
 567                                 SHA1_Update(&key->md,out,sha_off);
 568                                 aesni256_cbc_sha1_dec(in+aes_off,
 569                                         out+aes_off,blocks,&key->ks,ctx->iv,
 570                                         &key->md,out+sha_off);
 571
 572                                 sha_off += blocks*=SHA_CBLOCK;
 573                                 out += sha_off;
 574                                 len -= sha_off;
 575                                 inp_len -= sha_off;
 576
 577                                 key->md.Nl += (blocks<<3);      /* at most 18 bits */
 578                                 memcpy(ctx->iv,tail_iv,AES_BLOCK_SIZE);
 579                         }
 580 #endif
 581
 582 #if 1
 583                         len -= SHA_DIGEST_LENGTH;               /* amend mac */
 584                         if (len>=(256+SHA_CBLOCK)) {
 585                                 j = (len-(256+SHA_CBLOCK))&(0-SHA_CBLOCK);
 586                                 j += SHA_CBLOCK-key->md.num;
 587                                 SHA1_Update(&key->md,out,j);
 588                                 out += j;
 589                                 len -= j;
 590                                 inp_len -= j;
 591                         }
 592
 593                         /* but pretend as if we hashed padded payload */
 594                         bitlen = key->md.Nl+(inp_len<<3);       /* at most 18 bits */
 595 #ifdef BSWAP4
 596                         bitlen = BSWAP4(bitlen);
 597 #else
 598                         mac.c[0] = 0;
 599                         mac.c[1] = (unsigned char)(bitlen>>16);
 600                         mac.c[2] = (unsigned char)(bitlen>>8);
 601                         mac.c[3] = (unsigned char)bitlen;
 602                         bitlen = mac.u[0];
 603 #endif
 604
 605                         pmac->u[0]=0;
 606                         pmac->u[1]=0;
 607                         pmac->u[2]=0;
 608                         pmac->u[3]=0;
 609                         pmac->u[4]=0;
 610
 611                         for (res=key->md.num, j=0;j<len;j++) {
 612                                 size_t c = out[j];
 613                                 mask = (j-inp_len)>>(sizeof(j)*8-8);
 614                                 c &= mask;
 615                                 c |= 0x80&~mask&~((inp_len-j)>>(sizeof(j)*8-8));
 616                                 data->c[res++]=(unsigned char)c;
 617
 618                                 if (res!=SHA_CBLOCK) continue;
 619
 620                                 /* j is not incremented yet */
 621                                 mask = 0-((inp_len+7-j)>>(sizeof(j)*8-1));
 622                                 data->u[SHA_LBLOCK-1] |= bitlen&mask;
 623                                 sha1_block_data_order(&key->md,data,1);
 624                                 mask &= 0-((j-inp_len-72)>>(sizeof(j)*8-1));
 625                                 pmac->u[0] |= key->md.h0 & mask;
 626                                 pmac->u[1] |= key->md.h1 & mask;
 627                                 pmac->u[2] |= key->md.h2 & mask;
 628                                 pmac->u[3] |= key->md.h3 & mask;
 629                                 pmac->u[4] |= key->md.h4 & mask;
 630                                 res=0;
 631                         }
 632
 633                         for(i=res;i<SHA_CBLOCK;i++,j++) data->c[i]=0;
 634
 635                         if (res>SHA_CBLOCK-8) {
 636                                 mask = 0-((inp_len+8-j)>>(sizeof(j)*8-1));
 637                                 data->u[SHA_LBLOCK-1] |= bitlen&mask;
 638                                 sha1_block_data_order(&key->md,data,1);
 639                                 mask &= 0-((j-inp_len-73)>>(sizeof(j)*8-1));
 640                                 pmac->u[0] |= key->md.h0 & mask;
 641                                 pmac->u[1] |= key->md.h1 & mask;
 642                                 pmac->u[2] |= key->md.h2 & mask;
 643                                 pmac->u[3] |= key->md.h3 & mask;
 644                                 pmac->u[4] |= key->md.h4 & mask;
 645
 646                                 memset(data,0,SHA_CBLOCK);
 647                                 j+=64;
 648                         }
 649                         data->u[SHA_LBLOCK-1] = bitlen;
 650                         sha1_block_data_order(&key->md,data,1);
 651                         mask = 0-((j-inp_len-73)>>(sizeof(j)*8-1));
 652                         pmac->u[0] |= key->md.h0 & mask;
 653                         pmac->u[1] |= key->md.h1 & mask;
 654                         pmac->u[2] |= key->md.h2 & mask;
 655                         pmac->u[3] |= key->md.h3 & mask;
 656                         pmac->u[4] |= key->md.h4 & mask;
 657
 658 #ifdef BSWAP4
 659                         pmac->u[0] = BSWAP4(pmac->u[0]);
 660                         pmac->u[1] = BSWAP4(pmac->u[1]);
 661                         pmac->u[2] = BSWAP4(pmac->u[2]);
 662                         pmac->u[3] = BSWAP4(pmac->u[3]);
 663                         pmac->u[4] = BSWAP4(pmac->u[4]);
 664 #else
 665                         for (i=0;i<5;i++) {
 666                                 res = pmac->u[i];
 667                                 pmac->c[4*i+0]=(unsigned char)(res>>24);
 668                                 pmac->c[4*i+1]=(unsigned char)(res>>16);
 669                                 pmac->c[4*i+2]=(unsigned char)(res>>8);
 670                                 pmac->c[4*i+3]=(unsigned char)res;
 671                         }
 672 #endif
 673                         len += SHA_DIGEST_LENGTH;
 674 #else
 675                         SHA1_Update(&key->md,out,inp_len);
 676                         res = key->md.num;
 677                         SHA1_Final(pmac->c,&key->md);
 678
 679                         {
 680                         unsigned int inp_blocks, pad_blocks;
 681
 682                         /* but pretend as if we hashed padded payload */
 683                         inp_blocks = 1+((SHA_CBLOCK-9-res)>>(sizeof(res)*8-1));
 684                         res += (unsigned int)(len-inp_len);
 685                         pad_blocks = res / SHA_CBLOCK;
 686                         res %= SHA_CBLOCK;
 687                         pad_blocks += 1+((SHA_CBLOCK-9-res)>>(sizeof(res)*8-1));
 688                         for (;inp_blocks<pad_blocks;inp_blocks++)
 689                                 sha1_block_data_order(&key->md,data,1);
 690                         }
 691 #endif
 692                         key->md = key->tail;
 693                         SHA1_Update(&key->md,pmac->c,SHA_DIGEST_LENGTH);
 694                         SHA1_Final(pmac->c,&key->md);
 695
 696                         /* verify HMAC */
 697                         out += inp_len;
 698                         len -= inp_len;
 699 #if 1
 700                         {
 701                         unsigned char *p = out+len-1-maxpad-SHA_DIGEST_LENGTH;
 702                         size_t off = out-p;
 703                         unsigned int c, cmask;
 704
 705                         maxpad += SHA_DIGEST_LENGTH;
 706                         for (res=0,i=0,j=0;j<maxpad;j++) {
 707                                 c = p[j];
 708                                 cmask = ((int)(j-off-SHA_DIGEST_LENGTH))>>(sizeof(int)*8-1);
 709                                 res |= (c^pad)&~cmask;  /* ... and padding */
 710                                 cmask &= ((int)(off-1-j))>>(sizeof(int)*8-1);
 711                                 res |= (c^pmac->c[i])&cmask;
 712                                 i += 1&cmask;
 713                         }
 714                         maxpad -= SHA_DIGEST_LENGTH;
 715
 716                         res = 0-((0-res)>>(sizeof(res)*8-1));
 717                         ret &= (int)~res;
 718                         }
 719 #else
 720                         for (res=0,i=0;i<SHA_DIGEST_LENGTH;i++)
 721                                 res |= out[i]^pmac->c[i];
 722                         res = 0-((0-res)>>(sizeof(res)*8-1));
 723                         ret &= (int)~res;
 724
 725                         /* verify padding */
 726                         pad = (pad&~res) | (maxpad&res);
 727                         out = out+len-1-pad;
 728                         for (res=0,i=0;i<pad;i++)
 729                                 res |= out[i]^pad;
 730
 731                         res = (0-res)>>(sizeof(res)*8-1);
 732                         ret &= (int)~res;
 733 #endif
 734                         return ret;
 735                 } else {
 736 #if defined(STITCHED_DECRYPT_CALL)
 737                         if (len>=1024 && ctx->key_len==32) {
 738                                 if (sha_off%=SHA_CBLOCK)
 739                                         blocks = (len-3*SHA_CBLOCK)/SHA_CBLOCK;
 740                                 else
 741                                         blocks = (len-2*SHA_CBLOCK)/SHA_CBLOCK;
 742                                 aes_off = len-blocks*SHA_CBLOCK;
 743
 744                                 aesni_cbc_encrypt(in,out,aes_off,
 745                                         &key->ks,ctx->iv,0);
 746                                 SHA1_Update(&key->md,out,sha_off);
 747                                 aesni256_cbc_sha1_dec(in+aes_off,
 748                                         out+aes_off,blocks,&key->ks,ctx->iv,
 749                                         &key->md,out+sha_off);
 750
 751                                 sha_off += blocks*=SHA_CBLOCK;
 752                                 out += sha_off;
 753                                 len -= sha_off;
 754
 755                                 key->md.Nh += blocks>>29;
 756                                 key->md.Nl += blocks<<=3;
 757                                 if (key->md.Nl<(unsigned int)blocks) key->md.Nh++;
 758                         } else
 759 #endif
 760                         /* decrypt HMAC|padding at once */
 761                         aesni_cbc_encrypt(in,out,len,
 762                                         &key->ks,ctx->iv,0);
 763
 764                         SHA1_Update(&key->md,out,len);
 765                 }
 766         }
 767
 768         return 1;
 769         }
 770
 771 static int aesni_cbc_hmac_sha1_ctrl(EVP_CIPHER_CTX *ctx, int type, int arg, void *ptr)
 772         {
 773         EVP_AES_HMAC_SHA1 *key = data(ctx);
 774
 775         switch (type)
 776                 {
 777         case EVP_CTRL_AEAD_SET_MAC_KEY:
 778                 {
 779                 unsigned int  i;
 780                 unsigned char hmac_key[64];
 781
 782                 memset (hmac_key,0,sizeof(hmac_key));
 783
 784                 if (arg > (int)sizeof(hmac_key)) {
 785                         SHA1_Init(&key->head);
 786                         SHA1_Update(&key->head,ptr,arg);
 787                         SHA1_Final(hmac_key,&key->head);
 788                 } else {
 789                         memcpy(hmac_key,ptr,arg);
 790                 }
 791
 792                 for (i=0;i<sizeof(hmac_key);i++)
 793                         hmac_key[i] ^= 0x36;            /* ipad */
 794                 SHA1_Init(&key->head);
 795                 SHA1_Update(&key->head,hmac_key,sizeof(hmac_key));
 796
 797                 for (i=0;i<sizeof(hmac_key);i++)
 798                         hmac_key[i] ^= 0x36^0x5c;       /* opad */
 799                 SHA1_Init(&key->tail);
 800                 SHA1_Update(&key->tail,hmac_key,sizeof(hmac_key));
 801
 802                 OPENSSL_cleanse(hmac_key,sizeof(hmac_key));
 803
 804                 return 1;
 805                 }
 806         case EVP_CTRL_AEAD_TLS1_AAD:
 807                 {
 808                 unsigned char *p=ptr;
 809                 unsigned int   len=p[arg-2]<<8|p[arg-1];
 810
 811                 if (ctx->encrypt)
 812                         {
 813                         key->payload_length = len;
 814                         if ((key->aux.tls_ver=p[arg-4]<<8|p[arg-3]) >= TLS1_1_VERSION) {
 815                                 len -= AES_BLOCK_SIZE;
 816                                 p[arg-2] = len>>8;
 817                                 p[arg-1] = len;
 818                         }
 819                         key->md = key->head;
 820                         SHA1_Update(&key->md,p,arg);
 821
 822                         return (int)(((len+SHA_DIGEST_LENGTH+AES_BLOCK_SIZE)&-AES_BLOCK_SIZE)
 823                                 - len);
 824                         }
 825                 else
 826                         {
 827                         if (arg>13) arg = 13;
 828                         memcpy(key->aux.tls_aad,ptr,arg);
 829                         key->payload_length = arg;
 830
 831                         return SHA_DIGEST_LENGTH;
 832                         }
 833                 }
 834 #if !defined(OPENSSL_NO_MULTIBLOCK) && EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK
 835         case EVP_CTRL_TLS1_1_MULTIBLOCK_MAX_BUFSIZE:
 836                 return (int)(5+16+((arg+20+16)&-16));
 837         case EVP_CTRL_TLS1_1_MULTIBLOCK_AAD:
 838                 {
 839                 EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *param =
 840                         (EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *)ptr;
 841                 unsigned int n4x=1, x4;
 842                 unsigned int frag, last, packlen, inp_len;
 843
 844                 if (arg<(int)sizeof(EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM)) return -1;
 845
 846                 inp_len = param->inp[11]<<8|param->inp[12];
 847
 848                 if (ctx->encrypt)
 849                         {
 850                         if ((param->inp[9]<<8|param->inp[10]) < TLS1_1_VERSION)
 851                                 return -1;
 852
 853                         if (inp_len)
 854                                 {
 855                                 if (inp_len<4096) return 0;     /* too short */
 856
 857                                 if (inp_len>=8192 && OPENSSL_ia32cap_P[2]&(1<<5))
 858                                         n4x=2;  /* AVX2 */
 859                                 }
 860                         else if ((n4x=param->interleave/4) && n4x<=2)
 861                                 inp_len = param->len;
 862                         else
 863                                 return -1;
 864
 865                         key->md = key->head;
 866                         SHA1_Update(&key->md,param->inp,13);
 867
 868                         x4 = 4*n4x; n4x += 1;
 869
 870                         frag = inp_len>>n4x;
 871                         last = inp_len+frag-(frag<<n4x);
 872                         if (last>frag && ((last+13+9)%64<(x4-1))) {
 873                                 frag++;
 874                                 last -= x4-1;
 875                         }
 876
 877                         packlen = 5+16+((frag+20+16)&-16);
 878                         packlen = (packlen<<n4x)-packlen;
 879                         packlen += 5+16+((last+20+16)&-16);
 880
 881                         param->interleave = x4;
 882
 883                         return (int)packlen;
 884                         }
 885                 else
 886                         return -1;      /* not yet */
 887                 }
 888         case EVP_CTRL_TLS1_1_MULTIBLOCK_ENCRYPT:
 889                 {
 890                 EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *param =
 891                         (EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *)ptr;
 892
 893                 return (int)tls1_1_multi_block_encrypt(key,param->out,param->inp,
 894                                                 param->len,param->interleave/4);
 895                 }
 896         case EVP_CTRL_TLS1_1_MULTIBLOCK_DECRYPT:
 897 #endif
 898         default:
 899                 return -1;
 900                 }
 901         }
 902
 903 static EVP_CIPHER aesni_128_cbc_hmac_sha1_cipher =
 904         {
 905 #ifdef NID_aes_128_cbc_hmac_sha1
 906         NID_aes_128_cbc_hmac_sha1,
 907 #else
 908         NID_undef,
 909 #endif
 910         16,16,16,
 911         EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|
 912         EVP_CIPH_FLAG_AEAD_CIPHER|EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK,
 913         aesni_cbc_hmac_sha1_init_key,
 914         aesni_cbc_hmac_sha1_cipher,
 915         NULL,
 916         sizeof(EVP_AES_HMAC_SHA1),
 917         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_set_asn1_iv,
 918         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_get_asn1_iv,
 919         aesni_cbc_hmac_sha1_ctrl,
 920         NULL
 921         };
 922
 923 static EVP_CIPHER aesni_256_cbc_hmac_sha1_cipher =
 924         {
 925 #ifdef NID_aes_256_cbc_hmac_sha1
 926         NID_aes_256_cbc_hmac_sha1,
 927 #else
 928         NID_undef,
 929 #endif
 930         16,32,16,
 931         EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|
 932         EVP_CIPH_FLAG_AEAD_CIPHER|EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK,
 933         aesni_cbc_hmac_sha1_init_key,
 934         aesni_cbc_hmac_sha1_cipher,
 935         NULL,
 936         sizeof(EVP_AES_HMAC_SHA1),
 937         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_set_asn1_iv,
 938         EVP_CIPH_FLAG_DEFAULT_ASN1?NULL:EVP_CIPHER_get_asn1_iv,
 939         aesni_cbc_hmac_sha1_ctrl,
 940         NULL
 941         };
 942
 943 const EVP_CIPHER *EVP_aes_128_cbc_hmac_sha1(void)
 944         {
 945         return(OPENSSL_ia32cap_P[1]&AESNI_CAPABLE?
 946                 &aesni_128_cbc_hmac_sha1_cipher:NULL);
 947         }
 948
 949 const EVP_CIPHER *EVP_aes_256_cbc_hmac_sha1(void)
 950         {
 951         return(OPENSSL_ia32cap_P[1]&AESNI_CAPABLE?
 952                 &aesni_256_cbc_hmac_sha1_cipher:NULL);
 953         }
 954 #else
 955 const EVP_CIPHER *EVP_aes_128_cbc_hmac_sha1(void)
 956         {
 957         return NULL;
 958         }
 959 const EVP_CIPHER *EVP_aes_256_cbc_hmac_sha1(void)
 960         {
 961         return NULL;
 962         }
 963 #endif
 964 #endif