| 1 | /* |
| 2 | * Block matching used by the file-transfer code. |
| 3 | * |
| 4 | * Copyright (C) 1996 Andrew Tridgell |
| 5 | * Copyright (C) 1996 Paul Mackerras |
| 6 | * Copyright (C) 2003-2008 Wayne Davison |
| 7 | * |
| 8 | * This program is free software; you can redistribute it and/or modify |
| 9 | * it under the terms of the GNU General Public License as published by |
| 10 | * the Free Software Foundation; either version 3 of the License, or |
| 11 | * (at your option) any later version. |
| 12 | * |
| 13 | * This program is distributed in the hope that it will be useful, |
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 | * GNU General Public License for more details. |
| 17 | * |
| 18 | * You should have received a copy of the GNU General Public License along |
| 19 | * with this program; if not, visit the http://fsf.org website. |
| 20 | */ |
| 21 | |
| 22 | #include "rsync.h" |
| 23 | |
| 24 | extern int verbose; |
| 25 | extern int do_progress; |
| 26 | extern int checksum_seed; |
| 27 | extern int append_mode; |
| 28 | |
| 29 | int updating_basis_file; |
| 30 | |
| 31 | static int false_alarms; |
| 32 | static int hash_hits; |
| 33 | static int matches; |
| 34 | static int64 data_transfer; |
| 35 | |
| 36 | static int total_false_alarms; |
| 37 | static int total_hash_hits; |
| 38 | static int total_matches; |
| 39 | |
| 40 | extern struct stats stats; |
| 41 | |
| 42 | #define TRADITIONAL_TABLESIZE (1<<16) |
| 43 | |
| 44 | static uint32 tablesize; |
| 45 | static int32 *hash_table; |
| 46 | |
| 47 | #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF) |
| 48 | #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16) |
| 49 | |
| 50 | #define BIG_SUM2HASH(sum) ((sum)%tablesize) |
| 51 | |
| 52 | static void build_hash_table(struct sum_struct *s) |
| 53 | { |
| 54 | static uint32 alloc_size; |
| 55 | int32 i; |
| 56 | |
| 57 | /* Dynamically calculate the hash table size so that the hash load |
| 58 | * for big files is about 80%. A number greater than the traditional |
| 59 | * size must be odd or s2 will not be able to span the entire set. */ |
| 60 | tablesize = (uint32)(s->count/8) * 10 + 11; |
| 61 | if (tablesize < TRADITIONAL_TABLESIZE) |
| 62 | tablesize = TRADITIONAL_TABLESIZE; |
| 63 | if (tablesize > alloc_size || tablesize < alloc_size - 16*1024) { |
| 64 | if (hash_table) |
| 65 | free(hash_table); |
| 66 | hash_table = new_array(int32, tablesize); |
| 67 | if (!hash_table) |
| 68 | out_of_memory("build_hash_table"); |
| 69 | alloc_size = tablesize; |
| 70 | } |
| 71 | |
| 72 | memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]); |
| 73 | |
| 74 | if (tablesize == TRADITIONAL_TABLESIZE) { |
| 75 | for (i = 0; i < s->count; i++) { |
| 76 | uint32 t = SUM2HASH(s->sums[i].sum1); |
| 77 | s->sums[i].chain = hash_table[t]; |
| 78 | hash_table[t] = i; |
| 79 | } |
| 80 | } else { |
| 81 | for (i = 0; i < s->count; i++) { |
| 82 | uint32 t = BIG_SUM2HASH(s->sums[i].sum1); |
| 83 | s->sums[i].chain = hash_table[t]; |
| 84 | hash_table[t] = i; |
| 85 | } |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | |
| 90 | static OFF_T last_match; |
| 91 | |
| 92 | |
| 93 | /** |
| 94 | * Transmit a literal and/or match token. |
| 95 | * |
| 96 | * This delightfully-named function is called either when we find a |
| 97 | * match and need to transmit all the unmatched data leading up to it, |
| 98 | * or when we get bored of accumulating literal data and just need to |
| 99 | * transmit it. As a result of this second case, it is called even if |
| 100 | * we have not matched at all! |
| 101 | * |
| 102 | * @param i If >0, the number of a matched token. If 0, indicates we |
| 103 | * have only literal data. |
| 104 | **/ |
| 105 | static void matched(int f, struct sum_struct *s, struct map_struct *buf, |
| 106 | OFF_T offset, int32 i) |
| 107 | { |
| 108 | int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */ |
| 109 | int32 j; |
| 110 | |
| 111 | if (verbose > 2 && i >= 0) { |
| 112 | rprintf(FINFO, |
| 113 | "match at %.0f last_match=%.0f j=%d len=%ld n=%ld\n", |
| 114 | (double)offset, (double)last_match, i, |
| 115 | (long)s->sums[i].len, (long)n); |
| 116 | } |
| 117 | |
| 118 | send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len); |
| 119 | data_transfer += n; |
| 120 | |
| 121 | if (i >= 0) { |
| 122 | stats.matched_data += s->sums[i].len; |
| 123 | n += s->sums[i].len; |
| 124 | } |
| 125 | |
| 126 | for (j = 0; j < n; j += CHUNK_SIZE) { |
| 127 | int32 n1 = MIN(CHUNK_SIZE, n - j); |
| 128 | sum_update(map_ptr(buf, last_match + j, n1), n1); |
| 129 | } |
| 130 | |
| 131 | if (i >= 0) |
| 132 | last_match = offset + s->sums[i].len; |
| 133 | else |
| 134 | last_match = offset; |
| 135 | |
| 136 | if (buf && do_progress) |
| 137 | show_progress(last_match, buf->file_size); |
| 138 | } |
| 139 | |
| 140 | |
| 141 | static void hash_search(int f,struct sum_struct *s, |
| 142 | struct map_struct *buf, OFF_T len) |
| 143 | { |
| 144 | OFF_T offset, end; |
| 145 | int32 k, want_i, backup; |
| 146 | char sum2[SUM_LENGTH]; |
| 147 | uint32 s1, s2, sum; |
| 148 | int more; |
| 149 | schar *map; |
| 150 | |
| 151 | /* want_i is used to encourage adjacent matches, allowing the RLL |
| 152 | * coding of the output to work more efficiently. */ |
| 153 | want_i = 0; |
| 154 | |
| 155 | if (verbose > 2) { |
| 156 | rprintf(FINFO, "hash search b=%ld len=%.0f\n", |
| 157 | (long)s->blength, (double)len); |
| 158 | } |
| 159 | |
| 160 | k = (int32)MIN(len, (OFF_T)s->blength); |
| 161 | |
| 162 | map = (schar *)map_ptr(buf, 0, k); |
| 163 | |
| 164 | sum = get_checksum1((char *)map, k); |
| 165 | s1 = sum & 0xFFFF; |
| 166 | s2 = sum >> 16; |
| 167 | if (verbose > 3) |
| 168 | rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k); |
| 169 | |
| 170 | offset = 0; |
| 171 | |
| 172 | end = len + 1 - s->sums[s->count-1].len; |
| 173 | |
| 174 | if (verbose > 3) { |
| 175 | rprintf(FINFO, "hash search s->blength=%ld len=%.0f count=%.0f\n", |
| 176 | (long)s->blength, (double)len, (double)s->count); |
| 177 | } |
| 178 | |
| 179 | do { |
| 180 | int done_csum2 = 0; |
| 181 | int32 i; |
| 182 | |
| 183 | if (verbose > 4) { |
| 184 | rprintf(FINFO, "offset=%.0f sum=%04x%04x\n", |
| 185 | (double)offset, s2 & 0xFFFF, s1 & 0xFFFF); |
| 186 | } |
| 187 | |
| 188 | if (tablesize == TRADITIONAL_TABLESIZE) { |
| 189 | if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0) |
| 190 | goto null_hash; |
| 191 | sum = (s1 & 0xffff) | (s2 << 16); |
| 192 | } else { |
| 193 | sum = (s1 & 0xffff) | (s2 << 16); |
| 194 | if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0) |
| 195 | goto null_hash; |
| 196 | } |
| 197 | |
| 198 | hash_hits++; |
| 199 | do { |
| 200 | int32 l; |
| 201 | |
| 202 | if (sum != s->sums[i].sum1) |
| 203 | continue; |
| 204 | |
| 205 | /* also make sure the two blocks are the same length */ |
| 206 | l = (int32)MIN((OFF_T)s->blength, len-offset); |
| 207 | if (l != s->sums[i].len) |
| 208 | continue; |
| 209 | |
| 210 | /* in-place: ensure chunk's offset is either >= our |
| 211 | * offset or that the data didn't move. */ |
| 212 | if (updating_basis_file && s->sums[i].offset < offset |
| 213 | && !(s->sums[i].flags & SUMFLG_SAME_OFFSET)) |
| 214 | continue; |
| 215 | |
| 216 | if (verbose > 3) { |
| 217 | rprintf(FINFO, |
| 218 | "potential match at %.0f i=%ld sum=%08x\n", |
| 219 | (double)offset, (long)i, sum); |
| 220 | } |
| 221 | |
| 222 | if (!done_csum2) { |
| 223 | map = (schar *)map_ptr(buf,offset,l); |
| 224 | get_checksum2((char *)map,l,sum2); |
| 225 | done_csum2 = 1; |
| 226 | } |
| 227 | |
| 228 | if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) { |
| 229 | false_alarms++; |
| 230 | continue; |
| 231 | } |
| 232 | |
| 233 | /* When updating in-place, the best possible match is |
| 234 | * one with an identical offset, so we prefer that over |
| 235 | * the following want_i optimization. */ |
| 236 | if (updating_basis_file) { |
| 237 | int32 i2; |
| 238 | for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) { |
| 239 | if (s->sums[i2].offset != offset) |
| 240 | continue; |
| 241 | if (i2 != i) { |
| 242 | if (sum != s->sums[i2].sum1) |
| 243 | break; |
| 244 | if (memcmp(sum2, s->sums[i2].sum2, |
| 245 | s->s2length) != 0) |
| 246 | break; |
| 247 | i = i2; |
| 248 | } |
| 249 | /* This chunk was at the same offset on |
| 250 | * both the sender and the receiver. */ |
| 251 | s->sums[i].flags |= SUMFLG_SAME_OFFSET; |
| 252 | goto set_want_i; |
| 253 | } |
| 254 | } |
| 255 | |
| 256 | /* we've found a match, but now check to see |
| 257 | * if want_i can hint at a better match. */ |
| 258 | if (i != want_i && want_i < s->count |
| 259 | && (!updating_basis_file || s->sums[want_i].offset >= offset |
| 260 | || s->sums[want_i].flags & SUMFLG_SAME_OFFSET) |
| 261 | && sum == s->sums[want_i].sum1 |
| 262 | && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) { |
| 263 | /* we've found an adjacent match - the RLL coder |
| 264 | * will be happy */ |
| 265 | i = want_i; |
| 266 | } |
| 267 | set_want_i: |
| 268 | want_i = i + 1; |
| 269 | |
| 270 | matched(f,s,buf,offset,i); |
| 271 | offset += s->sums[i].len - 1; |
| 272 | k = (int32)MIN((OFF_T)s->blength, len-offset); |
| 273 | map = (schar *)map_ptr(buf, offset, k); |
| 274 | sum = get_checksum1((char *)map, k); |
| 275 | s1 = sum & 0xFFFF; |
| 276 | s2 = sum >> 16; |
| 277 | matches++; |
| 278 | break; |
| 279 | } while ((i = s->sums[i].chain) >= 0); |
| 280 | |
| 281 | null_hash: |
| 282 | backup = (int32)(offset - last_match); |
| 283 | /* We sometimes read 1 byte prior to last_match... */ |
| 284 | if (backup < 0) |
| 285 | backup = 0; |
| 286 | |
| 287 | /* Trim off the first byte from the checksum */ |
| 288 | more = offset + k < len; |
| 289 | map = (schar *)map_ptr(buf, offset - backup, k + more + backup) |
| 290 | + backup; |
| 291 | s1 -= map[0] + CHAR_OFFSET; |
| 292 | s2 -= k * (map[0]+CHAR_OFFSET); |
| 293 | |
| 294 | /* Add on the next byte (if there is one) to the checksum */ |
| 295 | if (more) { |
| 296 | s1 += map[k] + CHAR_OFFSET; |
| 297 | s2 += s1; |
| 298 | } else |
| 299 | --k; |
| 300 | |
| 301 | /* By matching early we avoid re-reading the |
| 302 | data 3 times in the case where a token |
| 303 | match comes a long way after last |
| 304 | match. The 3 reads are caused by the |
| 305 | running match, the checksum update and the |
| 306 | literal send. */ |
| 307 | if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE) |
| 308 | matched(f, s, buf, offset - s->blength, -2); |
| 309 | } while (++offset < end); |
| 310 | |
| 311 | matched(f, s, buf, len, -1); |
| 312 | map_ptr(buf, len-1, 1); |
| 313 | } |
| 314 | |
| 315 | |
| 316 | /** |
| 317 | * Scan through a origin file, looking for sections that match |
| 318 | * checksums from the generator, and transmit either literal or token |
| 319 | * data. |
| 320 | * |
| 321 | * Also calculates the MD4 checksum of the whole file, using the md |
| 322 | * accumulator. This is transmitted with the file as protection |
| 323 | * against corruption on the wire. |
| 324 | * |
| 325 | * @param s Checksums received from the generator. If <tt>s->count == |
| 326 | * 0</tt>, then there are actually no checksums for this file. |
| 327 | * |
| 328 | * @param len Length of the file to send. |
| 329 | **/ |
| 330 | void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len) |
| 331 | { |
| 332 | char file_sum[MAX_DIGEST_LEN]; |
| 333 | int sum_len; |
| 334 | |
| 335 | last_match = 0; |
| 336 | false_alarms = 0; |
| 337 | hash_hits = 0; |
| 338 | matches = 0; |
| 339 | data_transfer = 0; |
| 340 | |
| 341 | sum_init(checksum_seed); |
| 342 | |
| 343 | if (append_mode > 0) { |
| 344 | if (append_mode == 2) { |
| 345 | OFF_T j = 0; |
| 346 | for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) { |
| 347 | if (buf && do_progress) |
| 348 | show_progress(last_match, buf->file_size); |
| 349 | sum_update(map_ptr(buf, last_match, CHUNK_SIZE), |
| 350 | CHUNK_SIZE); |
| 351 | last_match = j; |
| 352 | } |
| 353 | if (last_match < s->flength) { |
| 354 | int32 n = (int32)(s->flength - last_match); |
| 355 | if (buf && do_progress) |
| 356 | show_progress(last_match, buf->file_size); |
| 357 | sum_update(map_ptr(buf, last_match, n), n); |
| 358 | } |
| 359 | } |
| 360 | last_match = s->flength; |
| 361 | s->count = 0; |
| 362 | } |
| 363 | |
| 364 | if (len > 0 && s->count > 0) { |
| 365 | build_hash_table(s); |
| 366 | |
| 367 | if (verbose > 2) |
| 368 | rprintf(FINFO,"built hash table\n"); |
| 369 | |
| 370 | hash_search(f, s, buf, len); |
| 371 | |
| 372 | if (verbose > 2) |
| 373 | rprintf(FINFO,"done hash search\n"); |
| 374 | } else { |
| 375 | OFF_T j; |
| 376 | /* by doing this in pieces we avoid too many seeks */ |
| 377 | for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE) |
| 378 | matched(f, s, buf, j, -2); |
| 379 | matched(f, s, buf, len, -1); |
| 380 | } |
| 381 | |
| 382 | sum_len = sum_end(file_sum); |
| 383 | /* If we had a read error, send a bad checksum. */ |
| 384 | if (buf && buf->status != 0) |
| 385 | file_sum[0]++; |
| 386 | |
| 387 | if (verbose > 2) |
| 388 | rprintf(FINFO,"sending file_sum\n"); |
| 389 | write_buf(f, file_sum, sum_len); |
| 390 | |
| 391 | if (verbose > 2) |
| 392 | rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n", |
| 393 | false_alarms, hash_hits, matches); |
| 394 | |
| 395 | total_hash_hits += hash_hits; |
| 396 | total_false_alarms += false_alarms; |
| 397 | total_matches += matches; |
| 398 | stats.literal_data += data_transfer; |
| 399 | } |
| 400 | |
| 401 | void match_report(void) |
| 402 | { |
| 403 | if (verbose <= 1) |
| 404 | return; |
| 405 | |
| 406 | rprintf(FINFO, |
| 407 | "total: matches=%d hash_hits=%d false_alarms=%d data=%.0f\n", |
| 408 | total_matches, total_hash_hits, total_false_alarms, |
| 409 | (double)stats.literal_data); |
| 410 | } |