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 | ||
58a1c1a2 | 94 | /* Transmit a literal and/or match token. |
f5f95a38 MP |
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 | * | |
58a1c1a2 WD |
102 | * If i >= 0, the number of a matched token. If < 0, indicates we have |
103 | * only literal data. A -1 will send a 0-token-int too, and a -2 sends | |
104 | * only literal data, w/o any token-int. */ | |
a06b419d | 105 | static void matched(int f, struct sum_struct *s, struct map_struct *buf, |
37a56445 | 106 | OFF_T offset, int32 i) |
c627d613 | 107 | { |
26404276 | 108 | int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */ |
a06b419d WD |
109 | int32 j; |
110 | ||
5a18b34d | 111 | if (DEBUG_GTE(DELTASUM, 2) && i >= 0) { |
a06b419d | 112 | rprintf(FINFO, |
6d56efa6 | 113 | "match at %s last_match=%s j=%d len=%ld n=%ld\n", |
adc2476f | 114 | big_num(offset), big_num(last_match), i, |
a06b419d WD |
115 | (long)s->sums[i].len, (long)n); |
116 | } | |
c627d613 | 117 | |
a06b419d | 118 | send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len); |
e7ebc36c | 119 | data_transfer += n; |
70d794dc | 120 | |
a800434a AT |
121 | if (i >= 0) { |
122 | stats.matched_data += s->sums[i].len; | |
e7ebc36c | 123 | n += s->sums[i].len; |
a800434a | 124 | } |
1e4f48d6 WD |
125 | |
126 | for (j = 0; j < n; j += CHUNK_SIZE) { | |
a06b419d WD |
127 | int32 n1 = MIN(CHUNK_SIZE, n - j); |
128 | sum_update(map_ptr(buf, last_match + j, n1), n1); | |
e7ebc36c | 129 | } |
9e31c482 | 130 | |
0d0e2e93 | 131 | if (i >= 0) |
e7ebc36c | 132 | last_match = offset + s->sums[i].len; |
0d0e2e93 AT |
133 | else |
134 | last_match = offset; | |
eb86d661 | 135 | |
951e826b | 136 | if (buf && INFO_GTE(PROGRESS, 1)) |
4440b8aa | 137 | show_progress(last_match, buf->file_size); |
9e31c482 AT |
138 | } |
139 | ||
140 | ||
c6e7fcb4 | 141 | static void hash_search(int f,struct sum_struct *s, |
da9d12f5 | 142 | struct map_struct *buf, OFF_T len) |
c627d613 | 143 | { |
51b2ff03 | 144 | OFF_T offset, aligned_offset, end; |
a2c770dc | 145 | int32 k, want_i, backup; |
e7ebc36c | 146 | char sum2[SUM_LENGTH]; |
1e4f48d6 | 147 | uint32 s1, s2, sum; |
7560c17a | 148 | int more; |
debb4505 | 149 | schar *map; |
e7ebc36c | 150 | |
a04d77bc WD |
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; | |
923fa978 | 154 | |
5a18b34d | 155 | if (DEBUG_GTE(DELTASUM, 2)) { |
6d56efa6 | 156 | rprintf(FINFO, "hash search b=%ld len=%s\n", |
adc2476f | 157 | (long)s->blength, big_num(len)); |
0e36d9da | 158 | } |
e7ebc36c | 159 | |
a06b419d | 160 | k = (int32)MIN(len, (OFF_T)s->blength); |
1e4f48d6 | 161 | |
0e36d9da | 162 | map = (schar *)map_ptr(buf, 0, k); |
1e4f48d6 | 163 | |
e7ebc36c AT |
164 | sum = get_checksum1((char *)map, k); |
165 | s1 = sum & 0xFFFF; | |
166 | s2 = sum >> 16; | |
5a18b34d | 167 | if (DEBUG_GTE(DELTASUM, 3)) |
a06b419d | 168 | rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k); |
1e4f48d6 | 169 | |
51b2ff03 | 170 | offset = aligned_offset = 0; |
1e4f48d6 | 171 | |
e7ebc36c | 172 | end = len + 1 - s->sums[s->count-1].len; |
1e4f48d6 | 173 | |
5a18b34d | 174 | if (DEBUG_GTE(DELTASUM, 3)) { |
6d56efa6 | 175 | rprintf(FINFO, "hash search s->blength=%ld len=%s count=%s\n", |
adc2476f | 176 | (long)s->blength, big_num(len), big_num(s->count)); |
0e36d9da | 177 | } |
1e4f48d6 | 178 | |
e7ebc36c | 179 | do { |
e7ebc36c | 180 | int done_csum2 = 0; |
d9bca0c3 | 181 | int32 i; |
1e4f48d6 | 182 | |
5a18b34d | 183 | if (DEBUG_GTE(DELTASUM, 4)) { |
6d56efa6 | 184 | rprintf(FINFO, "offset=%s sum=%04x%04x\n", |
adc2476f | 185 | big_num(offset), s2 & 0xFFFF, s1 & 0xFFFF); |
48cce779 | 186 | } |
1e4f48d6 | 187 | |
a2c770dc WD |
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 | } | |
900cfcb5 WD |
197 | |
198 | hash_hits++; | |
199 | do { | |
d9bca0c3 | 200 | int32 l; |
1e4f48d6 | 201 | |
0e36d9da WD |
202 | if (sum != s->sums[i].sum1) |
203 | continue; | |
1e4f48d6 | 204 | |
536541d5 | 205 | /* also make sure the two blocks are the same length */ |
a06b419d | 206 | l = (int32)MIN((OFF_T)s->blength, len-offset); |
0e36d9da WD |
207 | if (l != s->sums[i].len) |
208 | continue; | |
536541d5 | 209 | |
a36ffd39 | 210 | /* in-place: ensure chunk's offset is either >= our |
a3221d2a | 211 | * offset or that the data didn't move. */ |
a36ffd39 | 212 | if (updating_basis_file && s->sums[i].offset < offset |
a3221d2a WD |
213 | && !(s->sums[i].flags & SUMFLG_SAME_OFFSET)) |
214 | continue; | |
215 | ||
5a18b34d | 216 | if (DEBUG_GTE(DELTASUM, 3)) { |
d9bca0c3 | 217 | rprintf(FINFO, |
6d56efa6 | 218 | "potential match at %s i=%ld sum=%08x\n", |
adc2476f | 219 | big_num(offset), (long)i, sum); |
d9bca0c3 | 220 | } |
1e4f48d6 | 221 | |
e7ebc36c | 222 | if (!done_csum2) { |
debb4505 | 223 | map = (schar *)map_ptr(buf,offset,l); |
e7ebc36c AT |
224 | get_checksum2((char *)map,l,sum2); |
225 | done_csum2 = 1; | |
226 | } | |
1e4f48d6 | 227 | |
fc0257c9 | 228 | if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) { |
e7ebc36c AT |
229 | false_alarms++; |
230 | continue; | |
231 | } | |
923fa978 | 232 | |
a36ffd39 | 233 | /* When updating in-place, the best possible match is |
a3221d2a | 234 | * one with an identical offset, so we prefer that over |
51b2ff03 | 235 | * the adjacent want_i optimization. */ |
a36ffd39 | 236 | if (updating_basis_file) { |
8484ddd3 | 237 | /* All the generator's chunks start at blength boundaries. */ |
51b2ff03 WD |
238 | while (aligned_offset < offset) |
239 | aligned_offset += s->blength; | |
580cffde WD |
240 | if (offset == aligned_offset |
241 | || (sum == 0 && l == s->blength && aligned_offset + l <= len)) { | |
51b2ff03 WD |
242 | int32 i2; |
243 | for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) { | |
580cffde | 244 | if (s->sums[i2].offset != aligned_offset) |
51b2ff03 WD |
245 | continue; |
246 | if (i2 != i) { | |
247 | if (sum != s->sums[i2].sum1 | |
58a1c1a2 | 248 | || l != s->sums[i2].len |
51b2ff03 WD |
249 | || memcmp(sum2, s->sums[i2].sum2, s->s2length) != 0) |
250 | break; | |
251 | i = i2; | |
252 | } | |
580cffde WD |
253 | want_i = i; |
254 | if (offset != aligned_offset) { | |
255 | /* We've matched some zeros in a spot that is also zeros | |
256 | * further along in the basis file, if we find zeros ahead | |
257 | * in the sender's file, we'll output enough literal data | |
258 | * to re-align with the basis file, and get back to seeking | |
259 | * instead of writing. */ | |
3f26945c WD |
260 | backup = (int32)(aligned_offset - last_match); |
261 | if (backup < 0) | |
262 | backup = 0; | |
263 | map = (schar *)map_ptr(buf, aligned_offset - backup, l + backup) | |
264 | + backup; | |
580cffde WD |
265 | sum = get_checksum1((char *)map, l); |
266 | if (sum != s->sums[i2].sum1) | |
267 | break; | |
268 | get_checksum2((char *)map, l, sum2); | |
269 | if (memcmp(sum2, s->sums[i2].sum2, s->s2length) != 0) | |
270 | break; | |
271 | /* OK, we have a re-alignment match. Bump the offset | |
272 | * forward to the new match point. */ | |
273 | offset = aligned_offset; | |
274 | } | |
8484ddd3 | 275 | /* This chunk remained in the same spot in the old and new file. */ |
51b2ff03 | 276 | s->sums[i].flags |= SUMFLG_SAME_OFFSET; |
58a1c1a2 | 277 | break; |
a3221d2a | 278 | } |
d9bca0c3 | 279 | } |
a3221d2a WD |
280 | } |
281 | ||
923fa978 | 282 | /* we've found a match, but now check to see |
a04d77bc WD |
283 | * if want_i can hint at a better match. */ |
284 | if (i != want_i && want_i < s->count | |
a36ffd39 | 285 | && (!updating_basis_file || s->sums[want_i].offset >= offset |
a3221d2a | 286 | || s->sums[want_i].flags & SUMFLG_SAME_OFFSET) |
a04d77bc WD |
287 | && sum == s->sums[want_i].sum1 |
288 | && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) { | |
5e252dea WD |
289 | /* we've found an adjacent match - the RLL coder |
290 | * will be happy */ | |
a04d77bc | 291 | i = want_i; |
923fa978 | 292 | } |
a04d77bc | 293 | want_i = i + 1; |
1e4f48d6 | 294 | |
e7ebc36c AT |
295 | matched(f,s,buf,offset,i); |
296 | offset += s->sums[i].len - 1; | |
a06b419d | 297 | k = (int32)MIN((OFF_T)s->blength, len-offset); |
0e36d9da | 298 | map = (schar *)map_ptr(buf, offset, k); |
e7ebc36c AT |
299 | sum = get_checksum1((char *)map, k); |
300 | s1 = sum & 0xFFFF; | |
301 | s2 = sum >> 16; | |
302 | matches++; | |
303 | break; | |
900cfcb5 | 304 | } while ((i = s->sums[i].chain) >= 0); |
1e4f48d6 | 305 | |
900cfcb5 | 306 | null_hash: |
26404276 | 307 | backup = (int32)(offset - last_match); |
7560c17a WD |
308 | /* We sometimes read 1 byte prior to last_match... */ |
309 | if (backup < 0) | |
310 | backup = 0; | |
311 | ||
e7ebc36c | 312 | /* Trim off the first byte from the checksum */ |
7560c17a WD |
313 | more = offset + k < len; |
314 | map = (schar *)map_ptr(buf, offset - backup, k + more + backup) | |
315 | + backup; | |
e7ebc36c AT |
316 | s1 -= map[0] + CHAR_OFFSET; |
317 | s2 -= k * (map[0]+CHAR_OFFSET); | |
1e4f48d6 | 318 | |
e7ebc36c | 319 | /* Add on the next byte (if there is one) to the checksum */ |
7560c17a WD |
320 | if (more) { |
321 | s1 += map[k] + CHAR_OFFSET; | |
e7ebc36c | 322 | s2 += s1; |
0e36d9da | 323 | } else |
e7ebc36c | 324 | --k; |
45f133b9 | 325 | |
5914bf15 PM |
326 | /* By matching early we avoid re-reading the |
327 | data 3 times in the case where a token | |
328 | match comes a long way after last | |
329 | match. The 3 reads are caused by the | |
330 | running match, the checksum update and the | |
331 | literal send. */ | |
a06b419d WD |
332 | if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE) |
333 | matched(f, s, buf, offset - s->blength, -2); | |
e7ebc36c | 334 | } while (++offset < end); |
1e4f48d6 | 335 | |
a06b419d WD |
336 | matched(f, s, buf, len, -1); |
337 | map_ptr(buf, len-1, 1); | |
c627d613 AT |
338 | } |
339 | ||
340 | ||
f5f95a38 MP |
341 | /** |
342 | * Scan through a origin file, looking for sections that match | |
343 | * checksums from the generator, and transmit either literal or token | |
344 | * data. | |
345 | * | |
538ba24f MP |
346 | * Also calculates the MD4 checksum of the whole file, using the md |
347 | * accumulator. This is transmitted with the file as protection | |
348 | * against corruption on the wire. | |
349 | * | |
f5f95a38 MP |
350 | * @param s Checksums received from the generator. If <tt>s->count == |
351 | * 0</tt>, then there are actually no checksums for this file. | |
352 | * | |
353 | * @param len Length of the file to send. | |
354 | **/ | |
355 | void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len) | |
c627d613 | 356 | { |
e7ebc36c AT |
357 | last_match = 0; |
358 | false_alarms = 0; | |
900cfcb5 | 359 | hash_hits = 0; |
1e4f48d6 WD |
360 | matches = 0; |
361 | data_transfer = 0; | |
e7ebc36c | 362 | |
ba582f75 | 363 | sum_init(checksum_seed); |
e7ebc36c | 364 | |
edb97721 | 365 | if (append_mode > 0) { |
936fa865 WD |
366 | if (append_mode == 2) { |
367 | OFF_T j = 0; | |
368 | for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) { | |
951e826b | 369 | if (buf && INFO_GTE(PROGRESS, 1)) |
936fa865 WD |
370 | show_progress(last_match, buf->file_size); |
371 | sum_update(map_ptr(buf, last_match, CHUNK_SIZE), | |
372 | CHUNK_SIZE); | |
373 | last_match = j; | |
374 | } | |
375 | if (last_match < s->flength) { | |
376 | int32 n = (int32)(s->flength - last_match); | |
951e826b | 377 | if (buf && INFO_GTE(PROGRESS, 1)) |
936fa865 WD |
378 | show_progress(last_match, buf->file_size); |
379 | sum_update(map_ptr(buf, last_match, n), n); | |
380 | } | |
6cc11982 | 381 | } |
936fa865 | 382 | last_match = s->flength; |
6cc11982 WD |
383 | s->count = 0; |
384 | } | |
385 | ||
a06b419d | 386 | if (len > 0 && s->count > 0) { |
a2c770dc WD |
387 | build_hash_table(s); |
388 | ||
5a18b34d | 389 | if (DEBUG_GTE(DELTASUM, 2)) |
a2c770dc WD |
390 | rprintf(FINFO,"built hash table\n"); |
391 | ||
a0456b9c | 392 | hash_search(f, s, buf, len); |
1e4f48d6 | 393 | |
5a18b34d | 394 | if (DEBUG_GTE(DELTASUM, 2)) |
9486289c | 395 | rprintf(FINFO,"done hash search\n"); |
e7ebc36c | 396 | } else { |
eb86d661 AT |
397 | OFF_T j; |
398 | /* by doing this in pieces we avoid too many seeks */ | |
6cc11982 | 399 | for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE) |
a06b419d WD |
400 | matched(f, s, buf, j, -2); |
401 | matched(f, s, buf, len, -1); | |
e7ebc36c | 402 | } |
9e31c482 | 403 | |
886df221 WD |
404 | if (sum_end(sender_file_sum) != checksum_len) |
405 | overflow_exit("checksum_len"); /* Impossible... */ | |
406 | ||
407 | /* If we had a read error, send a bad checksum. We use all bits | |
408 | * off as long as the checksum doesn't happen to be that, in | |
409 | * which case we turn the last 0 bit into a 1. */ | |
410 | if (buf && buf->status != 0) { | |
411 | int i; | |
412 | for (i = 0; i < checksum_len && sender_file_sum[i] == 0; i++) {} | |
413 | memset(sender_file_sum, 0, checksum_len); | |
414 | if (i == checksum_len) | |
415 | sender_file_sum[i-1]++; | |
416 | } | |
c627d613 | 417 | |
5a18b34d | 418 | if (DEBUG_GTE(DELTASUM, 2)) |
bc63ae3f | 419 | rprintf(FINFO,"sending file_sum\n"); |
886df221 | 420 | write_buf(f, sender_file_sum, checksum_len); |
c627d613 | 421 | |
5a18b34d | 422 | if (DEBUG_GTE(DELTASUM, 2)) { |
900cfcb5 WD |
423 | rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n", |
424 | false_alarms, hash_hits, matches); | |
886df221 | 425 | } |
1e4f48d6 | 426 | |
900cfcb5 | 427 | total_hash_hits += hash_hits; |
e7ebc36c AT |
428 | total_false_alarms += false_alarms; |
429 | total_matches += matches; | |
a800434a | 430 | stats.literal_data += data_transfer; |
c627d613 AT |
431 | } |
432 | ||
433 | void match_report(void) | |
434 | { | |
5a18b34d | 435 | if (!DEBUG_GTE(DELTASUM, 1)) |
e7ebc36c | 436 | return; |
c627d613 | 437 | |
9486289c | 438 | rprintf(FINFO, |
6d56efa6 | 439 | "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n", |
900cfcb5 | 440 | total_matches, total_hash_hits, total_false_alarms, |
adc2476f | 441 | big_num(stats.literal_data)); |
c627d613 | 442 | } |