2 * Block matching used by the file-transfer code.
4 * Copyright (C) 1996 Andrew Tridgell
5 * Copyright (C) 1996 Paul Mackerras
6 * Copyright (C) 2003-2009 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 do_progress;
26 extern int checksum_seed;
27 extern int append_mode;
29 int updating_basis_file;
31 static int false_alarms;
34 static int64 data_transfer;
36 static int total_false_alarms;
37 static int total_hash_hits;
38 static int total_matches;
40 extern struct stats stats;
42 #define TRADITIONAL_TABLESIZE (1<<16)
44 static uint32 tablesize;
45 static int32 *hash_table;
47 #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF)
48 #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16)
50 #define BIG_SUM2HASH(sum) ((sum)%tablesize)
52 static void build_hash_table(struct sum_struct *s)
54 static uint32 alloc_size;
57 /* Dynamically calculate the hash table size so that the hash load
58 * for big files is about 80%. A number greater than the traditional
59 * size must be odd or s2 will not be able to span the entire set. */
60 tablesize = (uint32)(s->count/8) * 10 + 11;
61 if (tablesize < TRADITIONAL_TABLESIZE)
62 tablesize = TRADITIONAL_TABLESIZE;
63 if (tablesize > alloc_size || tablesize < alloc_size - 16*1024) {
66 hash_table = new_array(int32, tablesize);
68 out_of_memory("build_hash_table");
69 alloc_size = tablesize;
72 memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]);
74 if (tablesize == TRADITIONAL_TABLESIZE) {
75 for (i = 0; i < s->count; i++) {
76 uint32 t = SUM2HASH(s->sums[i].sum1);
77 s->sums[i].chain = hash_table[t];
81 for (i = 0; i < s->count; i++) {
82 uint32 t = BIG_SUM2HASH(s->sums[i].sum1);
83 s->sums[i].chain = hash_table[t];
90 static OFF_T last_match;
93 /* Transmit a literal and/or match token.
95 * This delightfully-named function is called either when we find a
96 * match and need to transmit all the unmatched data leading up to it,
97 * or when we get bored of accumulating literal data and just need to
98 * transmit it. As a result of this second case, it is called even if
99 * we have not matched at all!
101 * If i >= 0, the number of a matched token. If < 0, indicates we have
102 * only literal data. A -1 will send a 0-token-int too, and a -2 sends
103 * only literal data, w/o any token-int. */
104 static void matched(int f, struct sum_struct *s, struct map_struct *buf,
105 OFF_T offset, int32 i)
107 int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */
110 if (verbose > 2 && i >= 0) {
112 "match at %.0f last_match=%.0f j=%d len=%ld n=%ld\n",
113 (double)offset, (double)last_match, i,
114 (long)s->sums[i].len, (long)n);
117 send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len);
121 stats.matched_data += s->sums[i].len;
125 for (j = 0; j < n; j += CHUNK_SIZE) {
126 int32 n1 = MIN(CHUNK_SIZE, n - j);
127 sum_update(map_ptr(buf, last_match + j, n1), n1);
131 last_match = offset + s->sums[i].len;
135 if (buf && do_progress)
136 show_progress(last_match, buf->file_size);
140 static void hash_search(int f,struct sum_struct *s,
141 struct map_struct *buf, OFF_T len)
143 OFF_T offset, aligned_offset, end;
144 int32 k, want_i, backup;
145 char sum2[SUM_LENGTH];
150 /* want_i is used to encourage adjacent matches, allowing the RLL
151 * coding of the output to work more efficiently. */
155 rprintf(FINFO, "hash search b=%ld len=%.0f\n",
156 (long)s->blength, (double)len);
159 k = (int32)MIN(len, (OFF_T)s->blength);
161 map = (schar *)map_ptr(buf, 0, k);
163 sum = get_checksum1((char *)map, k);
167 rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
169 offset = aligned_offset = 0;
171 end = len + 1 - s->sums[s->count-1].len;
174 rprintf(FINFO, "hash search s->blength=%ld len=%.0f count=%.0f\n",
175 (long)s->blength, (double)len, (double)s->count);
183 rprintf(FINFO, "offset=%.0f sum=%04x%04x\n",
184 (double)offset, s2 & 0xFFFF, s1 & 0xFFFF);
187 if (tablesize == TRADITIONAL_TABLESIZE) {
188 if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0)
190 sum = (s1 & 0xffff) | (s2 << 16);
192 sum = (s1 & 0xffff) | (s2 << 16);
193 if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0)
201 if (sum != s->sums[i].sum1)
204 /* also make sure the two blocks are the same length */
205 l = (int32)MIN((OFF_T)s->blength, len-offset);
206 if (l != s->sums[i].len)
209 /* in-place: ensure chunk's offset is either >= our
210 * offset or that the data didn't move. */
211 if (updating_basis_file && s->sums[i].offset < offset
212 && !(s->sums[i].flags & SUMFLG_SAME_OFFSET))
217 "potential match at %.0f i=%ld sum=%08x\n",
218 (double)offset, (long)i, sum);
222 map = (schar *)map_ptr(buf,offset,l);
223 get_checksum2((char *)map,l,sum2);
227 if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) {
232 /* When updating in-place, the best possible match is
233 * one with an identical offset, so we prefer that over
234 * the adjacent want_i optimization. */
235 if (updating_basis_file) {
236 /* All the generator's chunks start at blength boundaries. */
237 while (aligned_offset < offset)
238 aligned_offset += s->blength;
239 if (offset == aligned_offset) {
241 for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) {
242 if (s->sums[i2].offset != offset)
245 if (sum != s->sums[i2].sum1
246 || l != s->sums[i2].len
247 || memcmp(sum2, s->sums[i2].sum2, s->s2length) != 0)
251 /* This chunk remained in the same spot in the old and new file. */
252 s->sums[i].flags |= SUMFLG_SAME_OFFSET;
259 /* we've found a match, but now check to see
260 * if want_i can hint at a better match. */
261 if (i != want_i && want_i < s->count
262 && (!updating_basis_file || s->sums[want_i].offset >= offset
263 || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
264 && sum == s->sums[want_i].sum1
265 && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
266 /* we've found an adjacent match - the RLL coder
272 matched(f,s,buf,offset,i);
273 offset += s->sums[i].len - 1;
274 k = (int32)MIN((OFF_T)s->blength, len-offset);
275 map = (schar *)map_ptr(buf, offset, k);
276 sum = get_checksum1((char *)map, k);
281 } while ((i = s->sums[i].chain) >= 0);
284 backup = (int32)(offset - last_match);
285 /* We sometimes read 1 byte prior to last_match... */
289 /* Trim off the first byte from the checksum */
290 more = offset + k < len;
291 map = (schar *)map_ptr(buf, offset - backup, k + more + backup)
293 s1 -= map[0] + CHAR_OFFSET;
294 s2 -= k * (map[0]+CHAR_OFFSET);
296 /* Add on the next byte (if there is one) to the checksum */
298 s1 += map[k] + CHAR_OFFSET;
303 /* By matching early we avoid re-reading the
304 data 3 times in the case where a token
305 match comes a long way after last
306 match. The 3 reads are caused by the
307 running match, the checksum update and the
309 if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
310 matched(f, s, buf, offset - s->blength, -2);
311 } while (++offset < end);
313 matched(f, s, buf, len, -1);
314 map_ptr(buf, len-1, 1);
319 * Scan through a origin file, looking for sections that match
320 * checksums from the generator, and transmit either literal or token
323 * Also calculates the MD4 checksum of the whole file, using the md
324 * accumulator. This is transmitted with the file as protection
325 * against corruption on the wire.
327 * @param s Checksums received from the generator. If <tt>s->count ==
328 * 0</tt>, then there are actually no checksums for this file.
330 * @param len Length of the file to send.
332 void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
334 char file_sum[MAX_DIGEST_LEN];
343 sum_init(checksum_seed);
345 if (append_mode > 0) {
346 if (append_mode == 2) {
348 for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
349 if (buf && do_progress)
350 show_progress(last_match, buf->file_size);
351 sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
355 if (last_match < s->flength) {
356 int32 n = (int32)(s->flength - last_match);
357 if (buf && do_progress)
358 show_progress(last_match, buf->file_size);
359 sum_update(map_ptr(buf, last_match, n), n);
362 last_match = s->flength;
366 if (len > 0 && s->count > 0) {
370 rprintf(FINFO,"built hash table\n");
372 hash_search(f, s, buf, len);
375 rprintf(FINFO,"done hash search\n");
378 /* by doing this in pieces we avoid too many seeks */
379 for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE)
380 matched(f, s, buf, j, -2);
381 matched(f, s, buf, len, -1);
384 sum_len = sum_end(file_sum);
385 /* If we had a read error, send a bad checksum. */
386 if (buf && buf->status != 0)
390 rprintf(FINFO,"sending file_sum\n");
391 write_buf(f, file_sum, sum_len);
394 rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
395 false_alarms, hash_hits, matches);
397 total_hash_hits += hash_hits;
398 total_false_alarms += false_alarms;
399 total_matches += matches;
400 stats.literal_data += data_transfer;
403 void match_report(void)
409 "total: matches=%d hash_hits=%d false_alarms=%d data=%.0f\n",
410 total_matches, total_hash_hits, total_false_alarms,
411 (double)stats.literal_data);