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) 2003-2013 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: #include "inums.h"
24:
25: extern int checksum_seed;
26: extern int append_mode;
27: extern int checksum_len;
28:
29: int updating_basis_file;
30: char sender_file_sum[MAX_DIGEST_LEN];
31:
32: static int false_alarms;
33: static int hash_hits;
34: static int matches;
35: static int64 data_transfer;
36:
37: static int total_false_alarms;
38: static int total_hash_hits;
39: static int total_matches;
40:
41: extern struct stats stats;
42:
43: #define TRADITIONAL_TABLESIZE (1<<16)
44:
45: static uint32 tablesize;
46: static int32 *hash_table;
47:
48: #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF)
49: #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16)
50:
51: #define BIG_SUM2HASH(sum) ((sum)%tablesize)
52:
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);
68: if (!hash_table)
69: out_of_memory("build_hash_table");
70: alloc_size = tablesize;
71: }
72:
73: memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]);
74:
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: }
87: }
88: }
89:
90:
91: static OFF_T last_match;
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: * 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. */
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 (DEBUG_GTE(DELTASUM, 2) && i >= 0) {
112: rprintf(FINFO,
113: "match at %s last_match=%s j=%d len=%ld n=%ld\n",
114: big_num(offset), big_num(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 && INFO_GTE(PROGRESS, 1))
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, aligned_offset, end;
145: int32 k, want_i, aligned_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 (DEBUG_GTE(DELTASUM, 2)) {
156: rprintf(FINFO, "hash search b=%ld len=%s\n",
157: (long)s->blength, big_num(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 (DEBUG_GTE(DELTASUM, 3))
168: rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
169:
170: offset = aligned_offset = aligned_i = 0;
171:
172: end = len + 1 - s->sums[s->count-1].len;
173:
174: if (DEBUG_GTE(DELTASUM, 3)) {
175: rprintf(FINFO, "hash search s->blength=%ld len=%s count=%s\n",
176: (long)s->blength, big_num(len), big_num(s->count));
177: }
178:
179: do {
180: int done_csum2 = 0;
181: uint32 hash_entry;
182: int32 i, *prev;
183:
184: if (DEBUG_GTE(DELTASUM, 4)) {
185: rprintf(FINFO, "offset=%s sum=%04x%04x\n",
186: big_num(offset), s2 & 0xFFFF, s1 & 0xFFFF);
187: }
188:
189: if (tablesize == TRADITIONAL_TABLESIZE) {
190: hash_entry = SUM2HASH2(s1,s2);
191: if ((i = hash_table[hash_entry]) < 0)
192: goto null_hash;
193: sum = (s1 & 0xffff) | (s2 << 16);
194: } else {
195: sum = (s1 & 0xffff) | (s2 << 16);
196: hash_entry = BIG_SUM2HASH(sum);
197: if ((i = hash_table[hash_entry]) < 0)
198: goto null_hash;
199: }
200: prev = &hash_table[hash_entry];
201:
202: hash_hits++;
203: do {
204: int32 l;
205:
206: /* When updating in-place, the chunk's offset must be
207: * either >= our offset or identical data at that offset.
208: * Remove any bypassed entries that we can never use. */
209: if (updating_basis_file && s->sums[i].offset < offset
210: && !(s->sums[i].flags & SUMFLG_SAME_OFFSET)) {
211: *prev = s->sums[i].chain;
212: continue;
213: }
214: prev = &s->sums[i].chain;
215:
216: if (sum != s->sums[i].sum1)
217: continue;
218:
219: /* also make sure the two blocks are the same length */
220: l = (int32)MIN((OFF_T)s->blength, len-offset);
221: if (l != s->sums[i].len)
222: continue;
223:
224: if (DEBUG_GTE(DELTASUM, 3)) {
225: rprintf(FINFO,
226: "potential match at %s i=%ld sum=%08x\n",
227: big_num(offset), (long)i, sum);
228: }
229:
230: if (!done_csum2) {
231: map = (schar *)map_ptr(buf,offset,l);
232: get_checksum2((char *)map,l,sum2);
233: done_csum2 = 1;
234: }
235:
236: if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) {
237: false_alarms++;
238: continue;
239: }
240:
241: /* When updating in-place, the best possible match is
242: * one with an identical offset, so we prefer that over
243: * the adjacent want_i optimization. */
244: if (updating_basis_file) {
245: /* All the generator's chunks start at blength boundaries. */
246: while (aligned_offset < offset) {
247: aligned_offset += s->blength;
248: aligned_i++;
249: }
250: if ((offset == aligned_offset
251: || (sum == 0 && l == s->blength && aligned_offset + l <= len))
252: && aligned_i < s->count) {
253: if (i != aligned_i) {
254: if (sum != s->sums[aligned_i].sum1
255: || l != s->sums[aligned_i].len
256: || memcmp(sum2, s->sums[aligned_i].sum2, s->s2length) != 0)
257: goto check_want_i;
258: i = aligned_i;
259: }
260: if (offset != aligned_offset) {
261: /* We've matched some zeros in a spot that is also zeros
262: * further along in the basis file, if we find zeros ahead
263: * in the sender's file, we'll output enough literal data
264: * to re-align with the basis file, and get back to seeking
265: * instead of writing. */
266: backup = (int32)(aligned_offset - last_match);
267: if (backup < 0)
268: backup = 0;
269: map = (schar *)map_ptr(buf, aligned_offset - backup, l + backup)
270: + backup;
271: sum = get_checksum1((char *)map, l);
272: if (sum != s->sums[i].sum1)
273: goto check_want_i;
274: get_checksum2((char *)map, l, sum2);
275: if (memcmp(sum2, s->sums[i].sum2, s->s2length) != 0)
276: goto check_want_i;
277: /* OK, we have a re-alignment match. Bump the offset
278: * forward to the new match point. */
279: offset = aligned_offset;
280: }
281: /* This identical chunk is in the same spot in the old and new file. */
282: s->sums[i].flags |= SUMFLG_SAME_OFFSET;
283: want_i = i;
284: }
285: }
286:
287: check_want_i:
288: /* we've found a match, but now check to see
289: * if want_i can hint at a better match. */
290: if (i != want_i && want_i < s->count
291: && (!updating_basis_file || s->sums[want_i].offset >= offset
292: || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
293: && sum == s->sums[want_i].sum1
294: && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
295: /* we've found an adjacent match - the RLL coder
296: * will be happy */
297: i = want_i;
298: }
299: want_i = i + 1;
300:
301: matched(f,s,buf,offset,i);
302: offset += s->sums[i].len - 1;
303: k = (int32)MIN((OFF_T)s->blength, len-offset);
304: map = (schar *)map_ptr(buf, offset, k);
305: sum = get_checksum1((char *)map, k);
306: s1 = sum & 0xFFFF;
307: s2 = sum >> 16;
308: matches++;
309: break;
310: } while ((i = s->sums[i].chain) >= 0);
311:
312: null_hash:
313: backup = (int32)(offset - last_match);
314: /* We sometimes read 1 byte prior to last_match... */
315: if (backup < 0)
316: backup = 0;
317:
318: /* Trim off the first byte from the checksum */
319: more = offset + k < len;
320: map = (schar *)map_ptr(buf, offset - backup, k + more + backup)
321: + backup;
322: s1 -= map[0] + CHAR_OFFSET;
323: s2 -= k * (map[0]+CHAR_OFFSET);
324:
325: /* Add on the next byte (if there is one) to the checksum */
326: if (more) {
327: s1 += map[k] + CHAR_OFFSET;
328: s2 += s1;
329: } else
330: --k;
331:
332: /* By matching early we avoid re-reading the
333: data 3 times in the case where a token
334: match comes a long way after last
335: match. The 3 reads are caused by the
336: running match, the checksum update and the
337: literal send. */
338: if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
339: matched(f, s, buf, offset - s->blength, -2);
340: } while (++offset < end);
341:
342: matched(f, s, buf, len, -1);
343: map_ptr(buf, len-1, 1);
344: }
345:
346:
347: /**
348: * Scan through a origin file, looking for sections that match
349: * checksums from the generator, and transmit either literal or token
350: * data.
351: *
352: * Also calculates the MD4 checksum of the whole file, using the md
353: * accumulator. This is transmitted with the file as protection
354: * against corruption on the wire.
355: *
356: * @param s Checksums received from the generator. If <tt>s->count ==
357: * 0</tt>, then there are actually no checksums for this file.
358: *
359: * @param len Length of the file to send.
360: **/
361: void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
362: {
363: last_match = 0;
364: false_alarms = 0;
365: hash_hits = 0;
366: matches = 0;
367: data_transfer = 0;
368:
369: sum_init(checksum_seed);
370:
371: if (append_mode > 0) {
372: if (append_mode == 2) {
373: OFF_T j = 0;
374: for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
375: if (buf && INFO_GTE(PROGRESS, 1))
376: show_progress(last_match, buf->file_size);
377: sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
378: CHUNK_SIZE);
379: last_match = j;
380: }
381: if (last_match < s->flength) {
382: int32 n = (int32)(s->flength - last_match);
383: if (buf && INFO_GTE(PROGRESS, 1))
384: show_progress(last_match, buf->file_size);
385: sum_update(map_ptr(buf, last_match, n), n);
386: }
387: }
388: last_match = s->flength;
389: s->count = 0;
390: }
391:
392: if (len > 0 && s->count > 0) {
393: build_hash_table(s);
394:
395: if (DEBUG_GTE(DELTASUM, 2))
396: rprintf(FINFO,"built hash table\n");
397:
398: hash_search(f, s, buf, len);
399:
400: if (DEBUG_GTE(DELTASUM, 2))
401: rprintf(FINFO,"done hash search\n");
402: } else {
403: OFF_T j;
404: /* by doing this in pieces we avoid too many seeks */
405: for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE)
406: matched(f, s, buf, j, -2);
407: matched(f, s, buf, len, -1);
408: }
409:
410: if (sum_end(sender_file_sum) != checksum_len)
411: overflow_exit("checksum_len"); /* Impossible... */
412:
413: /* If we had a read error, send a bad checksum. We use all bits
414: * off as long as the checksum doesn't happen to be that, in
415: * which case we turn the last 0 bit into a 1. */
416: if (buf && buf->status != 0) {
417: int i;
418: for (i = 0; i < checksum_len && sender_file_sum[i] == 0; i++) {}
419: memset(sender_file_sum, 0, checksum_len);
420: if (i == checksum_len)
421: sender_file_sum[i-1]++;
422: }
423:
424: if (DEBUG_GTE(DELTASUM, 2))
425: rprintf(FINFO,"sending file_sum\n");
426: write_buf(f, sender_file_sum, checksum_len);
427:
428: if (DEBUG_GTE(DELTASUM, 2)) {
429: rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
430: false_alarms, hash_hits, matches);
431: }
432:
433: total_hash_hits += hash_hits;
434: total_false_alarms += false_alarms;
435: total_matches += matches;
436: stats.literal_data += data_transfer;
437: }
438:
439: void match_report(void)
440: {
441: if (!DEBUG_GTE(DELTASUM, 1))
442: return;
443:
444: rprintf(FINFO,
445: "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n",
446: total_matches, total_hash_hits, total_false_alarms,
447: big_num(stats.literal_data));
448: }
FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>
![[BACK]](/icons/cvsweb/back.gif){kind=icon}
Return to match.c CVS log ![[TXT]](/icons/cvsweb/text.gif){kind=icon}
![[DIR]](/icons/cvsweb/dir.gif){kind=icon}
Up to [ELWIX - Embedded LightWeight unIX -] / embedaddon / rsync