2 * Block matching used by the file-transfer code.
4 * Copyright (C) 1996 Andrew Tridgell
5 * Copyright (C) 1996 Paul Mackerras
6 * Copyright (C) 2003-2020 Wayne Davison
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.
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.
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.
25 extern int checksum_seed;
26 extern int append_mode;
27 extern int xfersum_type;
29 int updating_basis_file;
30 char sender_file_sum[MAX_DIGEST_LEN];
32 static int false_alarms;
35 static int64 data_transfer;
37 static int total_false_alarms;
38 static int total_hash_hits;
39 static int total_matches;
41 extern struct stats stats;
43 #define TRADITIONAL_TABLESIZE (1<<16)
45 static uint32 tablesize;
46 static int32 *hash_table;
48 #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF)
49 #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16)
51 #define BIG_SUM2HASH(sum) ((sum)%tablesize)
53 static void build_hash_table(struct sum_struct *s)
55 static uint32 alloc_size;
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) {
67 hash_table = new_array(int32, tablesize);
68 alloc_size = tablesize;
71 memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]);
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];
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];
89 static OFF_T last_match;
92 /* Transmit a literal and/or match token.
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!
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)
105 int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */
108 if (DEBUG_GTE(DELTASUM, 2) && i >= 0) {
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);
115 send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len);
119 stats.matched_data += s->sums[i].len;
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);
129 last_match = offset + s->sums[i].len;
133 if (buf && INFO_GTE(PROGRESS, 1))
134 show_progress(last_match, buf->file_size);
138 static void hash_search(int f,struct sum_struct *s,
139 struct map_struct *buf, OFF_T len)
141 OFF_T offset, aligned_offset, end;
142 int32 k, want_i, aligned_i, backup;
143 char sum2[SUM_LENGTH];
148 /* want_i is used to encourage adjacent matches, allowing the RLL
149 * coding of the output to work more efficiently. */
152 if (DEBUG_GTE(DELTASUM, 2)) {
153 rprintf(FINFO, "hash search b=%ld len=%s\n",
154 (long)s->blength, big_num(len));
157 k = (int32)MIN(len, (OFF_T)s->blength);
159 map = (schar *)map_ptr(buf, 0, k);
161 sum = get_checksum1((char *)map, k);
164 if (DEBUG_GTE(DELTASUM, 3))
165 rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
167 offset = aligned_offset = aligned_i = 0;
169 end = len + 1 - s->sums[s->count-1].len;
171 if (DEBUG_GTE(DELTASUM, 3)) {
172 rprintf(FINFO, "hash search s->blength=%ld len=%s count=%s\n",
173 (long)s->blength, big_num(len), big_num(s->count));
181 if (DEBUG_GTE(DELTASUM, 4)) {
182 rprintf(FINFO, "offset=%s sum=%04x%04x\n",
183 big_num(offset), s2 & 0xFFFF, s1 & 0xFFFF);
186 if (tablesize == TRADITIONAL_TABLESIZE) {
187 hash_entry = SUM2HASH2(s1,s2);
188 if ((i = hash_table[hash_entry]) < 0)
190 sum = (s1 & 0xffff) | (s2 << 16);
192 sum = (s1 & 0xffff) | (s2 << 16);
193 hash_entry = BIG_SUM2HASH(sum);
194 if ((i = hash_table[hash_entry]) < 0)
197 prev = &hash_table[hash_entry];
203 /* When updating in-place, the chunk's offset must be
204 * either >= our offset or identical data at that offset.
205 * Remove any bypassed entries that we can never use. */
206 if (updating_basis_file && s->sums[i].offset < offset
207 && !(s->sums[i].flags & SUMFLG_SAME_OFFSET)) {
208 *prev = s->sums[i].chain;
211 prev = &s->sums[i].chain;
213 if (sum != s->sums[i].sum1)
216 /* also make sure the two blocks are the same length */
217 l = (int32)MIN((OFF_T)s->blength, len-offset);
218 if (l != s->sums[i].len)
221 if (DEBUG_GTE(DELTASUM, 3)) {
223 "potential match at %s i=%ld sum=%08x\n",
224 big_num(offset), (long)i, sum);
228 map = (schar *)map_ptr(buf,offset,l);
229 get_checksum2((char *)map,l,sum2);
233 if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) {
238 /* When updating in-place, the best possible match is
239 * one with an identical offset, so we prefer that over
240 * the adjacent want_i optimization. */
241 if (updating_basis_file) {
242 /* All the generator's chunks start at blength boundaries. */
243 while (aligned_offset < offset) {
244 aligned_offset += s->blength;
247 if ((offset == aligned_offset
248 || (sum == 0 && l == s->blength && aligned_offset + l <= len))
249 && aligned_i < s->count) {
250 if (i != aligned_i) {
251 if (sum != s->sums[aligned_i].sum1
252 || l != s->sums[aligned_i].len
253 || memcmp(sum2, s->sums[aligned_i].sum2, s->s2length) != 0)
257 if (offset != aligned_offset) {
258 /* We've matched some zeros in a spot that is also zeros
259 * further along in the basis file, if we find zeros ahead
260 * in the sender's file, we'll output enough literal data
261 * to re-align with the basis file, and get back to seeking
262 * instead of writing. */
263 backup = (int32)(aligned_offset - last_match);
266 map = (schar *)map_ptr(buf, aligned_offset - backup, l + backup)
268 sum = get_checksum1((char *)map, l);
269 if (sum != s->sums[i].sum1)
271 get_checksum2((char *)map, l, sum2);
272 if (memcmp(sum2, s->sums[i].sum2, s->s2length) != 0)
274 /* OK, we have a re-alignment match. Bump the offset
275 * forward to the new match point. */
276 offset = aligned_offset;
278 /* This identical chunk is in the same spot in the old and new file. */
279 s->sums[i].flags |= SUMFLG_SAME_OFFSET;
285 /* we've found a match, but now check to see
286 * if want_i can hint at a better match. */
287 if (i != want_i && want_i < s->count
288 && (!updating_basis_file || s->sums[want_i].offset >= offset
289 || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
290 && sum == s->sums[want_i].sum1
291 && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
292 /* we've found an adjacent match - the RLL coder
298 matched(f,s,buf,offset,i);
299 offset += s->sums[i].len - 1;
300 k = (int32)MIN((OFF_T)s->blength, len-offset);
301 map = (schar *)map_ptr(buf, offset, k);
302 sum = get_checksum1((char *)map, k);
307 } while ((i = s->sums[i].chain) >= 0);
310 backup = (int32)(offset - last_match);
311 /* We sometimes read 1 byte prior to last_match... */
315 /* Trim off the first byte from the checksum */
316 more = offset + k < len;
317 map = (schar *)map_ptr(buf, offset - backup, k + more + backup) + backup;
318 s1 -= map[0] + CHAR_OFFSET;
319 s2 -= k * (map[0]+CHAR_OFFSET);
321 /* Add on the next byte (if there is one) to the checksum */
323 s1 += map[k] + CHAR_OFFSET;
328 /* By matching early we avoid re-reading the
329 data 3 times in the case where a token
330 match comes a long way after last
331 match. The 3 reads are caused by the
332 running match, the checksum update and the
334 if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
335 matched(f, s, buf, offset - s->blength, -2);
336 } while (++offset < end);
338 matched(f, s, buf, len, -1);
339 map_ptr(buf, len-1, 1);
344 * Scan through a origin file, looking for sections that match
345 * checksums from the generator, and transmit either literal or token
348 * Also calculates the MD4 checksum of the whole file, using the md
349 * accumulator. This is transmitted with the file as protection
350 * against corruption on the wire.
352 * @param s Checksums received from the generator. If <tt>s->count ==
353 * 0</tt>, then there are actually no checksums for this file.
355 * @param len Length of the file to send.
357 void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
367 sum_init(xfersum_type, checksum_seed);
369 if (append_mode > 0) {
370 if (append_mode == 2) {
372 for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
373 if (buf && INFO_GTE(PROGRESS, 1))
374 show_progress(last_match, buf->file_size);
375 sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
379 if (last_match < s->flength) {
380 int32 n = (int32)(s->flength - last_match);
381 if (buf && INFO_GTE(PROGRESS, 1))
382 show_progress(last_match, buf->file_size);
383 sum_update(map_ptr(buf, last_match, n), n);
386 last_match = s->flength;
390 if (len > 0 && s->count > 0) {
393 if (DEBUG_GTE(DELTASUM, 2))
394 rprintf(FINFO,"built hash table\n");
396 hash_search(f, s, buf, len);
398 if (DEBUG_GTE(DELTASUM, 2))
399 rprintf(FINFO,"done hash search\n");
402 /* by doing this in pieces we avoid too many seeks */
403 for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE)
404 matched(f, s, buf, j, -2);
405 matched(f, s, buf, len, -1);
408 sum_len = sum_end(sender_file_sum);
410 /* If we had a read error, send a bad checksum. We use all bits
411 * off as long as the checksum doesn't happen to be that, in
412 * which case we turn the last 0 bit into a 1. */
413 if (buf && buf->status != 0) {
415 for (i = 0; i < sum_len && sender_file_sum[i] == 0; i++) {}
416 memset(sender_file_sum, 0, sum_len);
418 sender_file_sum[i-1]++;
421 if (DEBUG_GTE(DELTASUM, 2))
422 rprintf(FINFO,"sending file_sum\n");
423 write_buf(f, sender_file_sum, sum_len);
425 if (DEBUG_GTE(DELTASUM, 2)) {
426 rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
427 false_alarms, hash_hits, matches);
430 total_hash_hits += hash_hits;
431 total_false_alarms += false_alarms;
432 total_matches += matches;
433 stats.literal_data += data_transfer;
436 void match_report(void)
438 if (!DEBUG_GTE(DELTASUM, 1))
442 "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n",
443 total_matches, total_hash_hits, total_false_alarms,
444 big_num(stats.literal_data));