blake_large.c

   1 /* blake_large.c */
   2 /*
   3     This file is part of the AVR-Crypto-Lib.
   4     Copyright (C) 2009  Daniel Otte (daniel.otte@rub.de)
   5
   6     This program is free software: you can redistribute it and/or modify
   7     it under the terms of the GNU General Public License as published by
   8     the Free Software Foundation, either version 3 of the License, or
   9     (at your option) any later version.
  10
  11     This program is distributed in the hope that it will be useful,
  12     but WITHOUT ANY WARRANTY; without even the implied warranty of
  13     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14     GNU General Public License for more details.
  15
  16     You should have received a copy of the GNU General Public License
  17     along with this program.  If not, see <http://www.gnu.org/licenses/>.
  18 */
  19 /*
  20  * \file    blake_large.c
  21  * \author  Daniel Otte
  22  * \email   daniel.otte@rub.de
  23  * \date    2009-05-08
  24  * \license GPLv3 or later
  25  *
  26  */
  27
  28 #include <stdint.h>
  29 #include <string.h>
  30 #include <avr/pgmspace.h>
  31 #include "memxor.h"
  32 #include "blake_large.h"
  33 #include "blake_common.h"
  34
  35 #define BUG_3 0 /* bug compatibility with reference code */
  36 #define BUG_4 0 /* bug compatibility with reference code */
  37
  38 uint64_t pgm_read_qword(void* p){
  39         union{
  40                 uint64_t v64;
  41                 uint32_t v32[2];
  42         }r;
  43         r.v32[0] = pgm_read_dword(p);
  44         r.v32[1] = pgm_read_dword((uint8_t*)p+4);
  45         return r.v64;
  46 }
  47
  48 static
  49 uint64_t blake_c[] PROGMEM = {
  50    0x243F6A8885A308D3LL, 0x13198A2E03707344LL,
  51    0xA4093822299F31D0LL, 0x082EFA98EC4E6C89LL,
  52    0x452821E638D01377LL, 0xBE5466CF34E90C6CLL,
  53    0xC0AC29B7C97C50DDLL, 0x3F84D5B5B5470917LL,
  54    0x9216D5D98979FB1BLL, 0xD1310BA698DFB5ACLL,
  55    0x2FFD72DBD01ADFB7LL, 0xB8E1AFED6A267E96LL,
  56    0xBA7C9045F12C7F99LL, 0x24A19947B3916CF7LL,
  57    0x0801F2E2858EFC16LL, 0x636920D871574E69LL
  58
  59 };
  60
  61
  62
  63 #define ROTL64(a, n) (((a)<<(n))|((a)>>(64-(n))))
  64 #define ROTR64(a, n) (((a)>>(n))|((a)<<(64-(n))))
  65 #define CHANGE_ENDIAN32(a) (((a)<<24)| \
  66                             ((0x0000ff00&(a))<<8)| \
  67                                                     ((0x00ff0000&(a))>>8)| \
  68                                                     (a)>>24 )
  69
  70 void blake_large_g(uint8_t r, uint8_t i, uint64_t* v, const uint64_t* m){
  71         uint8_t a,b,c,d, s0, s1;
  72         a = pgm_read_byte(blake_index_lut+4*i+0);
  73         b = pgm_read_byte(blake_index_lut+4*i+1);
  74         c = pgm_read_byte(blake_index_lut+4*i+2);
  75         d = pgm_read_byte(blake_index_lut+4*i+3);
  76         s0 = pgm_read_byte(blake_sigma+16*r+2*i+0);
  77         s1 = pgm_read_byte(blake_sigma+16*r+2*i+1);
  78         v[a] += v[b] + (m[s0] ^ pgm_read_qword(&(blake_c[s1])));
  79         v[d]  = ROTR64(v[d]^v[a], 32);
  80         v[c] += v[d];
  81         v[b]  = ROTR64(v[b]^v[c], 25);
  82         v[a] += v[b] + (m[s1] ^ pgm_read_qword(&(blake_c[s0])));
  83         v[d]  = ROTR64(v[d]^v[a], 16);
  84         v[c] += v[d];
  85         v[b]  = ROTR64(v[b]^v[c], 11);
  86
  87 }
  88
  89 void blake_large_expand(uint64_t* v, const blake_large_ctx_t* ctx){
  90         uint8_t i;
  91         memcpy(v, ctx->h, 8*8);
  92         for(i=0; i<8; ++i){
  93                 v[8+i] = pgm_read_qword(&(blake_c[i]));
  94         }
  95         memxor((uint8_t*)v+8, ctx->s, 4*8);
  96
  97 }
  98
  99 void blake_large_changeendian(void* dest, const void* src){
 100         uint8_t i;
 101         uint32_t tmp;
 102         for(i=0; i<32; i+=2){
 103                 tmp = CHANGE_ENDIAN32(((uint32_t*)src)[i]);
 104                 ((uint32_t*)dest)[i] = CHANGE_ENDIAN32(((uint32_t*)src)[i+1]);
 105                 ((uint32_t*)dest)[i+1] = tmp;
 106         }
 107 }
 108
 109 void blake_large_compress(uint64_t* v,const void* m){
 110         uint8_t r,i;
 111         for(r=0; r<14; ++r){
 112                 for(i=0; i<8; ++i){
 113                         blake_large_g(r%10, i, v, (uint64_t*)m);
 114                 }
 115         }
 116 }
 117
 118 void blake_large_collapse(blake_large_ctx_t* ctx, uint64_t* v){
 119         uint8_t i;
 120         for(i=0; i<8; ++i){
 121                 ctx->h[i] ^= ctx->s[i%4] ^ v[i] ^ v[8+i];
 122         }
 123 }
 124
 125 void blake_large_nextBlock(blake_large_ctx_t* ctx, const void* msg){
 126         uint64_t v[16];
 127         uint64_t m[16];
 128         union {
 129                 uint64_t v64;
 130                 uint32_t v32[2];
 131         }ctr;
 132         blake_large_expand(v,ctx);
 133         ctx->counter++;
 134         ctr.v64 = ctx->counter*1024;
 135         v[12] ^= ctr.v64;
 136         v[13] ^= ctr.v64;
 137         blake_large_changeendian(m, msg);
 138         blake_large_compress(v, m);
 139         blake_large_collapse(ctx, v);
 140 }
 141
 142 void blake_large_lastBlock(blake_large_ctx_t* ctx, const void* msg, uint16_t length_b){
 143         while(length_b>=BLAKE_LARGE_BLOCKSIZE){
 144                 blake_large_nextBlock(ctx, msg);
 145                 msg = (uint8_t*)msg + BLAKE_LARGE_BLOCKSIZE_B;
 146                 length_b -= BLAKE_LARGE_BLOCKSIZE;
 147         }
 148         uint8_t buffer[128];
 149         uint64_t v[16];
 150         uint64_t ctr;
 151         ctr = ctx->counter*1024+length_b;
 152         memset(buffer, 0, 128);
 153         memcpy(buffer, msg, (length_b+7)/8);
 154         buffer[length_b/8] |= 0x80 >> (length_b%8);
 155         blake_large_changeendian(buffer, buffer);
 156         blake_large_expand(v, ctx);
 157 #if BUG_3
 158         uint8_t x=0;
 159         if(length_b%1024<895 && length_b%8)
 160                 x=0x40;
 161         v[12] ^= ctr + x;
 162         v[13] ^= ctr + x;
 163
 164 #else
 165         v[12] ^= ctr;
 166         v[13] ^= ctr;
 167 #endif
 168         if(length_b>1024-128-2){
 169 #if BUG_4
 170                 if(length_b<1017){
 171                         blake_large_compress(v, buffer);
 172                         blake_large_collapse(ctx, v);
 173                 }
 174 #else
 175                 blake_large_compress(v, buffer);
 176                 blake_large_collapse(ctx, v);
 177 #endif
 178                 memset(buffer, 0, 128-8);
 179                 blake_large_expand(v, ctx);
 180         }
 181         if(ctx->appendone)
 182                 buffer[128-16-8] |= 0x01;
 183         *((uint64_t*)(&(buffer[128-8]))) = ctr;
 184         blake_large_compress(v, buffer);
 185         blake_large_collapse(ctx, v);
 186
 187 }
 188
 189 uint64_t blake64_iv[] PROGMEM = {
 190     0x6A09E667F3BCC908LL, 0xBB67AE8584CAA73BLL,
 191     0x3C6EF372FE94F82BLL, 0xA54FF53A5F1D36F1LL,
 192     0x510E527FADE682D1LL, 0x9B05688C2B3E6C1FLL,
 193     0x1F83D9ABFB41BD6BLL, 0x5BE0CD19137E2179LL
 194 };
 195
 196 void blake64_init(blake64_ctx_t* ctx){
 197         uint8_t i;
 198         for(i=0; i<8; ++i){
 199                 ctx->h[i] = pgm_read_qword(&(blake64_iv[i]));
 200         }
 201         memset(ctx->s, 0, 4*8);
 202         ctx->counter = 0;
 203         ctx->appendone = 1;
 204 }
 205
 206 uint64_t blake48_iv[] PROGMEM = {
 207     0xCBBB9D5DC1059ED8LL, 0x629A292A367CD507LL,
 208     0x9159015A3070DD17LL, 0x152FECD8F70E5939LL,
 209     0x67332667FFC00B31LL, 0x8EB44A8768581511LL,
 210     0xDB0C2E0D64F98FA7LL, 0x47B5481DBEFA4FA4LL
 211 };
 212
 213 void blake48_init(blake48_ctx_t* ctx){
 214         uint8_t i;
 215         for(i=0; i<8; ++i){
 216                 ctx->h[i] = pgm_read_qword(&(blake48_iv[i]));
 217         }
 218         memset(ctx->s, 0, 4*8);
 219         ctx->counter = 0;
 220         ctx->appendone = 0;
 221 }
 222
 223 void blake64_ctx2hash(void* dest, const blake64_ctx_t* ctx){
 224         uint8_t i;
 225         for(i=0; i<8; ++i){
 226                 ((uint32_t*)dest)[2*i+0] = CHANGE_ENDIAN32((ctx->h[i])>>32);
 227                 ((uint32_t*)dest)[2*i+1] = CHANGE_ENDIAN32((uint32_t)ctx->h[i]);
 228         }
 229 }
 230
 231 void blake48_ctx2hash(void* dest, const blake48_ctx_t* ctx){
 232         uint8_t i;
 233         for(i=0; i<6; ++i){
 234                 ((uint32_t*)dest)[2*i+0] = CHANGE_ENDIAN32((ctx->h[i])>>32);
 235                 ((uint32_t*)dest)[2*i+1] = CHANGE_ENDIAN32((uint32_t)ctx->h[i]);
 236         }
 237 }
 238
 239 void blake64_nextBlock(blake64_ctx_t* ctx, const void* block){
 240         blake_large_nextBlock(ctx, block);
 241 }
 242
 243 void blake48_nextBlock(blake48_ctx_t* ctx, const void* block){
 244         blake_large_nextBlock(ctx, block);
 245 }
 246
 247 void blake64_lastBlock(blake64_ctx_t* ctx, const void* block, uint16_t length_b){
 248         blake_large_lastBlock(ctx, block, length_b);
 249 }
 250
 251 void blake48_lastBlock(blake48_ctx_t* ctx, const void* block, uint16_t length_b){
 252         blake_large_lastBlock(ctx, block, length_b);
 253 }
 254
 255 void blake64(void* dest, const void* msg, uint32_t length_b){
 256         blake_large_ctx_t ctx;
 257         blake64_init(&ctx);
 258         while(length_b>=BLAKE_LARGE_BLOCKSIZE){
 259                 blake_large_nextBlock(&ctx, msg);
 260                 msg = (uint8_t*)msg + BLAKE_LARGE_BLOCKSIZE_B;
 261                 length_b -= BLAKE_LARGE_BLOCKSIZE;
 262         }
 263         blake_large_lastBlock(&ctx, msg, length_b);
 264         blake64_ctx2hash(dest, &ctx);
 265 }
 266
 267 void blake48(void* dest, const void* msg, uint32_t length_b){
 268         blake_large_ctx_t ctx;
 269         blake48_init(&ctx);
 270         while(length_b>=BLAKE_LARGE_BLOCKSIZE){
 271                 blake_large_nextBlock(&ctx, msg);
 272                 msg = (uint8_t*)msg + BLAKE_LARGE_BLOCKSIZE_B;
 273                 length_b -= BLAKE_LARGE_BLOCKSIZE;
 274         }
 275         blake_large_lastBlock(&ctx, msg, length_b);
 276         blake48_ctx2hash(dest, &ctx);
 277 }