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) 2003-2020 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 xfersum_type;
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: alloc_size = tablesize;
69: }
70:
71: memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]);
72:
73: if (tablesize == TRADITIONAL_TABLESIZE) {
74: for (i = 0; i < s->count; i++) {
75: uint32 t = SUM2HASH(s->sums[i].sum1);
76: s->sums[i].chain = hash_table[t];
77: hash_table[t] = i;
78: }
79: } else {
80: for (i = 0; i < s->count; i++) {
81: uint32 t = BIG_SUM2HASH(s->sums[i].sum1);
82: s->sums[i].chain = hash_table[t];
83: hash_table[t] = i;
84: }
85: }
86: }
87:
88:
89: static OFF_T last_match;
90:
91:
92: /* Transmit a literal and/or match token.
93: *
94: * This delightfully-named function is called either when we find a
95: * match and need to transmit all the unmatched data leading up to it,
96: * or when we get bored of accumulating literal data and just need to
97: * transmit it. As a result of this second case, it is called even if
98: * we have not matched at all!
99: *
100: * If i >= 0, the number of a matched token. If < 0, indicates we have
101: * only literal data. A -1 will send a 0-token-int too, and a -2 sends
102: * only literal data, w/o any token-int. */
103: static void matched(int f, struct sum_struct *s, struct map_struct *buf, OFF_T offset, int32 i)
104: {
105: int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */
106: int32 j;
107:
108: if (DEBUG_GTE(DELTASUM, 2) && i >= 0) {
109: rprintf(FINFO,
110: "match at %s last_match=%s j=%d len=%ld n=%ld\n",
111: big_num(offset), big_num(last_match), i,
112: (long)s->sums[i].len, (long)n);
113: }
114:
115: send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len);
116: data_transfer += n;
117:
118: if (i >= 0) {
119: stats.matched_data += s->sums[i].len;
120: n += s->sums[i].len;
121: }
122:
123: for (j = 0; j < n; j += CHUNK_SIZE) {
124: int32 n1 = MIN(CHUNK_SIZE, n - j);
125: sum_update(map_ptr(buf, last_match + j, n1), n1);
126: }
127:
128: if (i >= 0)
129: last_match = offset + s->sums[i].len;
130: else
131: last_match = offset;
132:
133: if (buf && INFO_GTE(PROGRESS, 1))
134: show_progress(last_match, buf->file_size);
135: }
136:
137:
138: static void hash_search(int f,struct sum_struct *s,
139: struct map_struct *buf, OFF_T len)
140: {
141: OFF_T offset, aligned_offset, end;
142: int32 k, want_i, aligned_i, backup;
143: char sum2[SUM_LENGTH];
144: uint32 s1, s2, sum;
145: int more;
146: schar *map;
147:
148: /* want_i is used to encourage adjacent matches, allowing the RLL
149: * coding of the output to work more efficiently. */
150: want_i = 0;
151:
152: if (DEBUG_GTE(DELTASUM, 2)) {
153: rprintf(FINFO, "hash search b=%ld len=%s\n",
154: (long)s->blength, big_num(len));
155: }
156:
157: k = (int32)MIN(len, (OFF_T)s->blength);
158:
159: map = (schar *)map_ptr(buf, 0, k);
160:
161: sum = get_checksum1((char *)map, k);
162: s1 = sum & 0xFFFF;
163: s2 = sum >> 16;
164: if (DEBUG_GTE(DELTASUM, 3))
165: rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
166:
167: checksum2_enable_prefetch(buf, len, s->blength);
168:
169: offset = aligned_offset = aligned_i = 0;
170:
171: end = len + 1 - s->sums[s->count-1].len;
172:
173: if (DEBUG_GTE(DELTASUM, 3)) {
174: rprintf(FINFO, "hash search s->blength=%ld len=%s count=%s\n",
175: (long)s->blength, big_num(len), big_num(s->count));
176: }
177:
178: do {
179: int done_csum2 = 0;
180: uint32 hash_entry;
181: int32 i, *prev;
182:
183: if (DEBUG_GTE(DELTASUM, 4)) {
184: rprintf(FINFO, "offset=%s sum=%04x%04x\n",
185: big_num(offset), s2 & 0xFFFF, s1 & 0xFFFF);
186: }
187:
188: if (tablesize == TRADITIONAL_TABLESIZE) {
189: hash_entry = SUM2HASH2(s1,s2);
190: if ((i = hash_table[hash_entry]) < 0)
191: goto null_hash;
192: sum = (s1 & 0xffff) | (s2 << 16);
193: } else {
194: sum = (s1 & 0xffff) | (s2 << 16);
195: hash_entry = BIG_SUM2HASH(sum);
196: if ((i = hash_table[hash_entry]) < 0)
197: goto null_hash;
198: }
199: prev = &hash_table[hash_entry];
200:
201: hash_hits++;
202: do {
203: int32 l;
204:
205: /* When updating in-place, the chunk's offset must be
206: * either >= our offset or identical data at that offset.
207: * Remove any bypassed entries that we can never use. */
208: if (updating_basis_file && s->sums[i].offset < offset
209: && !(s->sums[i].flags & SUMFLG_SAME_OFFSET)) {
210: *prev = s->sums[i].chain;
211: continue;
212: }
213: prev = &s->sums[i].chain;
214:
215: if (sum != s->sums[i].sum1)
216: continue;
217:
218: /* also make sure the two blocks are the same length */
219: l = (int32)MIN((OFF_T)s->blength, len-offset);
220: if (l != s->sums[i].len)
221: continue;
222:
223: if (DEBUG_GTE(DELTASUM, 3)) {
224: rprintf(FINFO,
225: "potential match at %s i=%ld sum=%08x\n",
226: big_num(offset), (long)i, sum);
227: }
228:
229: if (!done_csum2) {
230: map = (schar *)map_ptr(buf,offset,l);
231: get_checksum2((char *)map, l, sum2, offset);
232: done_csum2 = 1;
233: }
234:
235: if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) {
236: false_alarms++;
237: continue;
238: }
239:
240: /* When updating in-place, the best possible match is
241: * one with an identical offset, so we prefer that over
242: * the adjacent want_i optimization. */
243: if (updating_basis_file) {
244: /* All the generator's chunks start at blength boundaries. */
245: while (aligned_offset < offset) {
246: aligned_offset += s->blength;
247: aligned_i++;
248: }
249: if ((offset == aligned_offset
250: || (sum == 0 && l == s->blength && aligned_offset + l <= len))
251: && aligned_i < s->count) {
252: if (i != aligned_i) {
253: if (sum != s->sums[aligned_i].sum1
254: || l != s->sums[aligned_i].len
255: || memcmp(sum2, s->sums[aligned_i].sum2, s->s2length) != 0)
256: goto check_want_i;
257: i = aligned_i;
258: }
259: if (offset != aligned_offset) {
260: /* We've matched some zeros in a spot that is also zeros
261: * further along in the basis file, if we find zeros ahead
262: * in the sender's file, we'll output enough literal data
263: * to re-align with the basis file, and get back to seeking
264: * instead of writing. */
265: backup = (int32)(aligned_offset - last_match);
266: if (backup < 0)
267: backup = 0;
268: map = (schar *)map_ptr(buf, aligned_offset - backup, l + backup)
269: + backup;
270: sum = get_checksum1((char *)map, l);
271: if (sum != s->sums[i].sum1)
272: goto check_want_i;
273: get_checksum2((char *)map, l, sum2, aligned_offset);
274: if (memcmp(sum2, s->sums[i].sum2, s->s2length) != 0)
275: goto check_want_i;
276: /* OK, we have a re-alignment match. Bump the offset
277: * forward to the new match point. */
278: offset = aligned_offset;
279: }
280: /* This identical chunk is in the same spot in the old and new file. */
281: s->sums[i].flags |= SUMFLG_SAME_OFFSET;
282: want_i = i;
283: }
284: }
285:
286: check_want_i:
287: /* we've found a match, but now check to see
288: * if want_i can hint at a better match. */
289: if (i != want_i && want_i < s->count
290: && (!updating_basis_file || s->sums[want_i].offset >= offset
291: || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
292: && sum == s->sums[want_i].sum1
293: && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
294: /* we've found an adjacent match - the RLL coder
295: * will be happy */
296: i = want_i;
297: }
298: want_i = i + 1;
299:
300: matched(f,s,buf,offset,i);
301: offset += s->sums[i].len - 1;
302: k = (int32)MIN((OFF_T)s->blength, len-offset);
303: map = (schar *)map_ptr(buf, offset, k);
304: sum = get_checksum1((char *)map, k);
305: s1 = sum & 0xFFFF;
306: s2 = sum >> 16;
307: matches++;
308: break;
309: } while ((i = s->sums[i].chain) >= 0);
310:
311: null_hash:
312: backup = (int32)(offset - last_match);
313: /* We sometimes read 1 byte prior to last_match... */
314: if (backup < 0)
315: backup = 0;
316:
317: /* Trim off the first byte from the checksum */
318: more = offset + k < len;
319: map = (schar *)map_ptr(buf, offset - backup, k + more + backup) + backup;
320: s1 -= map[0] + CHAR_OFFSET;
321: s2 -= k * (map[0]+CHAR_OFFSET);
322:
323: /* Add on the next byte (if there is one) to the checksum */
324: if (more) {
325: s1 += map[k] + CHAR_OFFSET;
326: s2 += s1;
327: } else
328: --k;
329:
330: /* By matching early we avoid re-reading the
331: data 3 times in the case where a token
332: match comes a long way after last
333: match. The 3 reads are caused by the
334: running match, the checksum update and the
335: literal send. */
336: if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
337: matched(f, s, buf, offset - s->blength, -2);
338: } while (++offset < end);
339:
340: checksum2_disable_prefetch();
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: int sum_len;
364:
365: last_match = 0;
366: false_alarms = 0;
367: hash_hits = 0;
368: matches = 0;
369: data_transfer = 0;
370:
371: sum_init(xfersum_type, checksum_seed);
372:
373: if (append_mode > 0) {
374: if (append_mode == 2) {
375: OFF_T j = 0;
376: for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
377: if (buf && INFO_GTE(PROGRESS, 1))
378: show_progress(last_match, buf->file_size);
379: sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
380: CHUNK_SIZE);
381: last_match = j;
382: }
383: if (last_match < s->flength) {
384: int32 n = (int32)(s->flength - last_match);
385: if (buf && INFO_GTE(PROGRESS, 1))
386: show_progress(last_match, buf->file_size);
387: sum_update(map_ptr(buf, last_match, n), n);
388: }
389: }
390: last_match = s->flength;
391: s->count = 0;
392: }
393:
394: if (len > 0 && s->count > 0) {
395: build_hash_table(s);
396:
397: if (DEBUG_GTE(DELTASUM, 2))
398: rprintf(FINFO,"built hash table\n");
399:
400: hash_search(f, s, buf, len);
401:
402: if (DEBUG_GTE(DELTASUM, 2))
403: rprintf(FINFO,"done hash search\n");
404: } else {
405: OFF_T j;
406: /* by doing this in pieces we avoid too many seeks */
407: for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE)
408: matched(f, s, buf, j, -2);
409: matched(f, s, buf, len, -1);
410: }
411:
412: sum_len = sum_end(sender_file_sum);
413:
414: /* If we had a read error, send a bad checksum. We use all bits
415: * off as long as the checksum doesn't happen to be that, in
416: * which case we turn the last 0 bit into a 1. */
417: if (buf && buf->status != 0) {
418: int i;
419: for (i = 0; i < sum_len && sender_file_sum[i] == 0; i++) {}
420: memset(sender_file_sum, 0, sum_len);
421: if (i == sum_len)
422: sender_file_sum[i-1]++;
423: }
424:
425: if (DEBUG_GTE(DELTASUM, 2))
426: rprintf(FINFO,"sending file_sum\n");
427: write_buf(f, sender_file_sum, sum_len);
428:
429: if (DEBUG_GTE(DELTASUM, 2)) {
430: rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
431: false_alarms, hash_hits, matches);
432: }
433:
434: total_hash_hits += hash_hits;
435: total_false_alarms += false_alarms;
436: total_matches += matches;
437: stats.literal_data += data_transfer;
438: }
439:
440: void match_report(void)
441: {
442: if (!DEBUG_GTE(DELTASUM, 1))
443: return;
444:
445: rprintf(FINFO,
446: "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n",
447: total_matches, total_hash_hits, total_false_alarms,
448: big_num(stats.literal_data));
449: }
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