Commit | Line | Data |
---|---|---|
1e4f48d6 | 1 | /* |
0f78b815 WD |
2 | * Block matching used by the file-transfer code. |
3 | * | |
4 | * Copyright (C) 1996 Andrew Tridgell | |
5 | * Copyright (C) 1996 Paul Mackerras | |
b3bf9b9d | 6 | * Copyright (C) 2003-2009 Wayne Davison |
0f78b815 WD |
7 | * |
8 | * This program is free software; you can redistribute it and/or modify | |
8e41b68e WD |
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. | |
0f78b815 WD |
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 | * | |
e7c67065 | 18 | * You should have received a copy of the GNU General Public License along |
4fd842f9 | 19 | * with this program; if not, visit the http://fsf.org website. |
0f78b815 | 20 | */ |
c627d613 AT |
21 | |
22 | #include "rsync.h" | |
5dd14f0c | 23 | #include "inums.h" |
c627d613 | 24 | |
ba582f75 | 25 | extern int checksum_seed; |
6cc11982 | 26 | extern int append_mode; |
886df221 | 27 | extern int checksum_len; |
a36ffd39 WD |
28 | |
29 | int updating_basis_file; | |
886df221 | 30 | char sender_file_sum[MAX_DIGEST_LEN]; |
c627d613 | 31 | |
c627d613 | 32 | static int false_alarms; |
900cfcb5 | 33 | static int hash_hits; |
c627d613 | 34 | static int matches; |
a800434a | 35 | static int64 data_transfer; |
c627d613 | 36 | |
3a6a366f | 37 | static int total_false_alarms; |
900cfcb5 | 38 | static int total_hash_hits; |
3a6a366f | 39 | static int total_matches; |
c627d613 | 40 | |
a800434a | 41 | extern struct stats stats; |
c627d613 | 42 | |
a2c770dc | 43 | #define TRADITIONAL_TABLESIZE (1<<16) |
48cce779 | 44 | |
a2c770dc | 45 | static uint32 tablesize; |
68207537 | 46 | static int32 *hash_table; |
c627d613 | 47 | |
48cce779 WD |
48 | #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF) |
49 | #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16) | |
c627d613 | 50 | |
a2c770dc | 51 | #define BIG_SUM2HASH(sum) ((sum)%tablesize) |
48cce779 | 52 | |
a2c770dc WD |
53 | static void build_hash_table(struct sum_struct *s) |
54 | { | |
55 | static uint32 alloc_size; | |
56 | int32 i; | |
57 | ||
58 | /* Dynamically calculate the hash table size so that the hash load | |
59 | * for big files is about 80%. A number greater than the traditional | |
60 | * size must be odd or s2 will not be able to span the entire set. */ | |
61 | tablesize = (uint32)(s->count/8) * 10 + 11; | |
62 | if (tablesize < TRADITIONAL_TABLESIZE) | |
63 | tablesize = TRADITIONAL_TABLESIZE; | |
64 | if (tablesize > alloc_size || tablesize < alloc_size - 16*1024) { | |
65 | if (hash_table) | |
66 | free(hash_table); | |
67 | hash_table = new_array(int32, tablesize); | |
68207537 | 68 | if (!hash_table) |
d9bca0c3 | 69 | out_of_memory("build_hash_table"); |
a2c770dc | 70 | alloc_size = tablesize; |
d9bca0c3 | 71 | } |
c627d613 | 72 | |
a2c770dc | 73 | memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]); |
4c17cdcb | 74 | |
a2c770dc WD |
75 | if (tablesize == TRADITIONAL_TABLESIZE) { |
76 | for (i = 0; i < s->count; i++) { | |
77 | uint32 t = SUM2HASH(s->sums[i].sum1); | |
78 | s->sums[i].chain = hash_table[t]; | |
79 | hash_table[t] = i; | |
80 | } | |
81 | } else { | |
82 | for (i = 0; i < s->count; i++) { | |
83 | uint32 t = BIG_SUM2HASH(s->sums[i].sum1); | |
84 | s->sums[i].chain = hash_table[t]; | |
85 | hash_table[t] = i; | |
86 | } | |
4c17cdcb | 87 | } |
c627d613 AT |
88 | } |
89 | ||
90 | ||
bcacc18b | 91 | static OFF_T last_match; |
c627d613 AT |
92 | |
93 | ||
f5f95a38 MP |
94 | /** |
95 | * Transmit a literal and/or match token. | |
96 | * | |
97 | * This delightfully-named function is called either when we find a | |
98 | * match and need to transmit all the unmatched data leading up to it, | |
99 | * or when we get bored of accumulating literal data and just need to | |
100 | * transmit it. As a result of this second case, it is called even if | |
101 | * we have not matched at all! | |
102 | * | |
103 | * @param i If >0, the number of a matched token. If 0, indicates we | |
104 | * have only literal data. | |
105 | **/ | |
a06b419d | 106 | static void matched(int f, struct sum_struct *s, struct map_struct *buf, |
37a56445 | 107 | OFF_T offset, int32 i) |
c627d613 | 108 | { |
26404276 | 109 | int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */ |
a06b419d WD |
110 | int32 j; |
111 | ||
5a18b34d | 112 | if (DEBUG_GTE(DELTASUM, 2) && i >= 0) { |
a06b419d | 113 | rprintf(FINFO, |
6d56efa6 | 114 | "match at %s last_match=%s j=%d len=%ld n=%ld\n", |
adc2476f | 115 | big_num(offset), big_num(last_match), i, |
a06b419d WD |
116 | (long)s->sums[i].len, (long)n); |
117 | } | |
c627d613 | 118 | |
a06b419d | 119 | send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len); |
e7ebc36c | 120 | data_transfer += n; |
70d794dc | 121 | |
a800434a AT |
122 | if (i >= 0) { |
123 | stats.matched_data += s->sums[i].len; | |
e7ebc36c | 124 | n += s->sums[i].len; |
a800434a | 125 | } |
1e4f48d6 WD |
126 | |
127 | for (j = 0; j < n; j += CHUNK_SIZE) { | |
a06b419d WD |
128 | int32 n1 = MIN(CHUNK_SIZE, n - j); |
129 | sum_update(map_ptr(buf, last_match + j, n1), n1); | |
e7ebc36c | 130 | } |
9e31c482 | 131 | |
0d0e2e93 | 132 | if (i >= 0) |
e7ebc36c | 133 | last_match = offset + s->sums[i].len; |
0d0e2e93 AT |
134 | else |
135 | last_match = offset; | |
eb86d661 | 136 | |
951e826b | 137 | if (buf && INFO_GTE(PROGRESS, 1)) |
4440b8aa | 138 | show_progress(last_match, buf->file_size); |
9e31c482 AT |
139 | } |
140 | ||
141 | ||
c6e7fcb4 | 142 | static void hash_search(int f,struct sum_struct *s, |
da9d12f5 | 143 | struct map_struct *buf, OFF_T len) |
c627d613 | 144 | { |
a2c770dc WD |
145 | OFF_T offset, end; |
146 | int32 k, want_i, backup; | |
e7ebc36c | 147 | char sum2[SUM_LENGTH]; |
1e4f48d6 | 148 | uint32 s1, s2, sum; |
7560c17a | 149 | int more; |
debb4505 | 150 | schar *map; |
e7ebc36c | 151 | |
a04d77bc WD |
152 | /* want_i is used to encourage adjacent matches, allowing the RLL |
153 | * coding of the output to work more efficiently. */ | |
154 | want_i = 0; | |
923fa978 | 155 | |
5a18b34d | 156 | if (DEBUG_GTE(DELTASUM, 2)) { |
6d56efa6 | 157 | rprintf(FINFO, "hash search b=%ld len=%s\n", |
adc2476f | 158 | (long)s->blength, big_num(len)); |
0e36d9da | 159 | } |
e7ebc36c | 160 | |
a06b419d | 161 | k = (int32)MIN(len, (OFF_T)s->blength); |
1e4f48d6 | 162 | |
0e36d9da | 163 | map = (schar *)map_ptr(buf, 0, k); |
1e4f48d6 | 164 | |
e7ebc36c AT |
165 | sum = get_checksum1((char *)map, k); |
166 | s1 = sum & 0xFFFF; | |
167 | s2 = sum >> 16; | |
5a18b34d | 168 | if (DEBUG_GTE(DELTASUM, 3)) |
a06b419d | 169 | rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k); |
1e4f48d6 | 170 | |
e7ebc36c | 171 | offset = 0; |
1e4f48d6 | 172 | |
e7ebc36c | 173 | end = len + 1 - s->sums[s->count-1].len; |
1e4f48d6 | 174 | |
5a18b34d | 175 | if (DEBUG_GTE(DELTASUM, 3)) { |
6d56efa6 | 176 | rprintf(FINFO, "hash search s->blength=%ld len=%s count=%s\n", |
adc2476f | 177 | (long)s->blength, big_num(len), big_num(s->count)); |
0e36d9da | 178 | } |
1e4f48d6 | 179 | |
e7ebc36c | 180 | do { |
e7ebc36c | 181 | int done_csum2 = 0; |
d9bca0c3 | 182 | int32 i; |
1e4f48d6 | 183 | |
5a18b34d | 184 | if (DEBUG_GTE(DELTASUM, 4)) { |
6d56efa6 | 185 | rprintf(FINFO, "offset=%s sum=%04x%04x\n", |
adc2476f | 186 | big_num(offset), s2 & 0xFFFF, s1 & 0xFFFF); |
48cce779 | 187 | } |
1e4f48d6 | 188 | |
a2c770dc WD |
189 | if (tablesize == TRADITIONAL_TABLESIZE) { |
190 | if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0) | |
191 | goto null_hash; | |
192 | sum = (s1 & 0xffff) | (s2 << 16); | |
193 | } else { | |
194 | sum = (s1 & 0xffff) | (s2 << 16); | |
195 | if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0) | |
196 | goto null_hash; | |
197 | } | |
900cfcb5 WD |
198 | |
199 | hash_hits++; | |
200 | do { | |
d9bca0c3 | 201 | int32 l; |
1e4f48d6 | 202 | |
0e36d9da WD |
203 | if (sum != s->sums[i].sum1) |
204 | continue; | |
1e4f48d6 | 205 | |
536541d5 | 206 | /* also make sure the two blocks are the same length */ |
a06b419d | 207 | l = (int32)MIN((OFF_T)s->blength, len-offset); |
0e36d9da WD |
208 | if (l != s->sums[i].len) |
209 | continue; | |
536541d5 | 210 | |
a36ffd39 | 211 | /* in-place: ensure chunk's offset is either >= our |
a3221d2a | 212 | * offset or that the data didn't move. */ |
a36ffd39 | 213 | if (updating_basis_file && s->sums[i].offset < offset |
a3221d2a WD |
214 | && !(s->sums[i].flags & SUMFLG_SAME_OFFSET)) |
215 | continue; | |
216 | ||
5a18b34d | 217 | if (DEBUG_GTE(DELTASUM, 3)) { |
d9bca0c3 | 218 | rprintf(FINFO, |
6d56efa6 | 219 | "potential match at %s i=%ld sum=%08x\n", |
adc2476f | 220 | big_num(offset), (long)i, sum); |
d9bca0c3 | 221 | } |
1e4f48d6 | 222 | |
e7ebc36c | 223 | if (!done_csum2) { |
debb4505 | 224 | map = (schar *)map_ptr(buf,offset,l); |
e7ebc36c AT |
225 | get_checksum2((char *)map,l,sum2); |
226 | done_csum2 = 1; | |
227 | } | |
1e4f48d6 | 228 | |
fc0257c9 | 229 | if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) { |
e7ebc36c AT |
230 | false_alarms++; |
231 | continue; | |
232 | } | |
923fa978 | 233 | |
a36ffd39 | 234 | /* When updating in-place, the best possible match is |
a3221d2a WD |
235 | * one with an identical offset, so we prefer that over |
236 | * the following want_i optimization. */ | |
a36ffd39 | 237 | if (updating_basis_file) { |
d9bca0c3 WD |
238 | int32 i2; |
239 | for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) { | |
a3221d2a WD |
240 | if (s->sums[i2].offset != offset) |
241 | continue; | |
242 | if (i2 != i) { | |
243 | if (sum != s->sums[i2].sum1) | |
244 | break; | |
245 | if (memcmp(sum2, s->sums[i2].sum2, | |
246 | s->s2length) != 0) | |
247 | break; | |
248 | i = i2; | |
249 | } | |
250 | /* This chunk was at the same offset on | |
251 | * both the sender and the receiver. */ | |
252 | s->sums[i].flags |= SUMFLG_SAME_OFFSET; | |
253 | goto set_want_i; | |
d9bca0c3 | 254 | } |
a3221d2a WD |
255 | } |
256 | ||
923fa978 | 257 | /* we've found a match, but now check to see |
a04d77bc WD |
258 | * if want_i can hint at a better match. */ |
259 | if (i != want_i && want_i < s->count | |
a36ffd39 | 260 | && (!updating_basis_file || s->sums[want_i].offset >= offset |
a3221d2a | 261 | || s->sums[want_i].flags & SUMFLG_SAME_OFFSET) |
a04d77bc WD |
262 | && sum == s->sums[want_i].sum1 |
263 | && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) { | |
5e252dea WD |
264 | /* we've found an adjacent match - the RLL coder |
265 | * will be happy */ | |
a04d77bc | 266 | i = want_i; |
923fa978 | 267 | } |
a3221d2a | 268 | set_want_i: |
a04d77bc | 269 | want_i = i + 1; |
1e4f48d6 | 270 | |
e7ebc36c AT |
271 | matched(f,s,buf,offset,i); |
272 | offset += s->sums[i].len - 1; | |
a06b419d | 273 | k = (int32)MIN((OFF_T)s->blength, len-offset); |
0e36d9da | 274 | map = (schar *)map_ptr(buf, offset, k); |
e7ebc36c AT |
275 | sum = get_checksum1((char *)map, k); |
276 | s1 = sum & 0xFFFF; | |
277 | s2 = sum >> 16; | |
278 | matches++; | |
279 | break; | |
900cfcb5 | 280 | } while ((i = s->sums[i].chain) >= 0); |
1e4f48d6 | 281 | |
900cfcb5 | 282 | null_hash: |
26404276 | 283 | backup = (int32)(offset - last_match); |
7560c17a WD |
284 | /* We sometimes read 1 byte prior to last_match... */ |
285 | if (backup < 0) | |
286 | backup = 0; | |
287 | ||
e7ebc36c | 288 | /* Trim off the first byte from the checksum */ |
7560c17a WD |
289 | more = offset + k < len; |
290 | map = (schar *)map_ptr(buf, offset - backup, k + more + backup) | |
291 | + backup; | |
e7ebc36c AT |
292 | s1 -= map[0] + CHAR_OFFSET; |
293 | s2 -= k * (map[0]+CHAR_OFFSET); | |
1e4f48d6 | 294 | |
e7ebc36c | 295 | /* Add on the next byte (if there is one) to the checksum */ |
7560c17a WD |
296 | if (more) { |
297 | s1 += map[k] + CHAR_OFFSET; | |
e7ebc36c | 298 | s2 += s1; |
0e36d9da | 299 | } else |
e7ebc36c | 300 | --k; |
45f133b9 | 301 | |
5914bf15 PM |
302 | /* By matching early we avoid re-reading the |
303 | data 3 times in the case where a token | |
304 | match comes a long way after last | |
305 | match. The 3 reads are caused by the | |
306 | running match, the checksum update and the | |
307 | literal send. */ | |
a06b419d WD |
308 | if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE) |
309 | matched(f, s, buf, offset - s->blength, -2); | |
e7ebc36c | 310 | } while (++offset < end); |
1e4f48d6 | 311 | |
a06b419d WD |
312 | matched(f, s, buf, len, -1); |
313 | map_ptr(buf, len-1, 1); | |
c627d613 AT |
314 | } |
315 | ||
316 | ||
f5f95a38 MP |
317 | /** |
318 | * Scan through a origin file, looking for sections that match | |
319 | * checksums from the generator, and transmit either literal or token | |
320 | * data. | |
321 | * | |
538ba24f MP |
322 | * Also calculates the MD4 checksum of the whole file, using the md |
323 | * accumulator. This is transmitted with the file as protection | |
324 | * against corruption on the wire. | |
325 | * | |
f5f95a38 MP |
326 | * @param s Checksums received from the generator. If <tt>s->count == |
327 | * 0</tt>, then there are actually no checksums for this file. | |
328 | * | |
329 | * @param len Length of the file to send. | |
330 | **/ | |
331 | void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len) | |
c627d613 | 332 | { |
e7ebc36c AT |
333 | last_match = 0; |
334 | false_alarms = 0; | |
900cfcb5 | 335 | hash_hits = 0; |
1e4f48d6 WD |
336 | matches = 0; |
337 | data_transfer = 0; | |
e7ebc36c | 338 | |
ba582f75 | 339 | sum_init(checksum_seed); |
e7ebc36c | 340 | |
edb97721 | 341 | if (append_mode > 0) { |
936fa865 WD |
342 | if (append_mode == 2) { |
343 | OFF_T j = 0; | |
344 | for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) { | |
951e826b | 345 | if (buf && INFO_GTE(PROGRESS, 1)) |
936fa865 WD |
346 | show_progress(last_match, buf->file_size); |
347 | sum_update(map_ptr(buf, last_match, CHUNK_SIZE), | |
348 | CHUNK_SIZE); | |
349 | last_match = j; | |
350 | } | |
351 | if (last_match < s->flength) { | |
352 | int32 n = (int32)(s->flength - last_match); | |
951e826b | 353 | if (buf && INFO_GTE(PROGRESS, 1)) |
936fa865 WD |
354 | show_progress(last_match, buf->file_size); |
355 | sum_update(map_ptr(buf, last_match, n), n); | |
356 | } | |
6cc11982 | 357 | } |
936fa865 | 358 | last_match = s->flength; |
6cc11982 WD |
359 | s->count = 0; |
360 | } | |
361 | ||
a06b419d | 362 | if (len > 0 && s->count > 0) { |
a2c770dc WD |
363 | build_hash_table(s); |
364 | ||
5a18b34d | 365 | if (DEBUG_GTE(DELTASUM, 2)) |
a2c770dc WD |
366 | rprintf(FINFO,"built hash table\n"); |
367 | ||
a0456b9c | 368 | hash_search(f, s, buf, len); |
1e4f48d6 | 369 | |
5a18b34d | 370 | if (DEBUG_GTE(DELTASUM, 2)) |
9486289c | 371 | rprintf(FINFO,"done hash search\n"); |
e7ebc36c | 372 | } else { |
eb86d661 AT |
373 | OFF_T j; |
374 | /* by doing this in pieces we avoid too many seeks */ | |
6cc11982 | 375 | for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE) |
a06b419d WD |
376 | matched(f, s, buf, j, -2); |
377 | matched(f, s, buf, len, -1); | |
e7ebc36c | 378 | } |
9e31c482 | 379 | |
886df221 WD |
380 | if (sum_end(sender_file_sum) != checksum_len) |
381 | overflow_exit("checksum_len"); /* Impossible... */ | |
382 | ||
383 | /* If we had a read error, send a bad checksum. We use all bits | |
384 | * off as long as the checksum doesn't happen to be that, in | |
385 | * which case we turn the last 0 bit into a 1. */ | |
386 | if (buf && buf->status != 0) { | |
387 | int i; | |
388 | for (i = 0; i < checksum_len && sender_file_sum[i] == 0; i++) {} | |
389 | memset(sender_file_sum, 0, checksum_len); | |
390 | if (i == checksum_len) | |
391 | sender_file_sum[i-1]++; | |
392 | } | |
c627d613 | 393 | |
5a18b34d | 394 | if (DEBUG_GTE(DELTASUM, 2)) |
bc63ae3f | 395 | rprintf(FINFO,"sending file_sum\n"); |
886df221 | 396 | write_buf(f, sender_file_sum, checksum_len); |
c627d613 | 397 | |
5a18b34d | 398 | if (DEBUG_GTE(DELTASUM, 2)) { |
900cfcb5 WD |
399 | rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n", |
400 | false_alarms, hash_hits, matches); | |
886df221 | 401 | } |
1e4f48d6 | 402 | |
900cfcb5 | 403 | total_hash_hits += hash_hits; |
e7ebc36c AT |
404 | total_false_alarms += false_alarms; |
405 | total_matches += matches; | |
a800434a | 406 | stats.literal_data += data_transfer; |
c627d613 AT |
407 | } |
408 | ||
409 | void match_report(void) | |
410 | { | |
5a18b34d | 411 | if (!DEBUG_GTE(DELTASUM, 1)) |
e7ebc36c | 412 | return; |
c627d613 | 413 | |
9486289c | 414 | rprintf(FINFO, |
6d56efa6 | 415 | "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n", |
900cfcb5 | 416 | total_matches, total_hash_hits, total_false_alarms, |
adc2476f | 417 | big_num(stats.literal_data)); |
c627d613 | 418 | } |