2 * Block matching used by the file-transfer code.
4 * Copyright (C) 1996 Andrew Tridgell
5 * Copyright (C) 1996 Paul Mackerras
6 * Copyright (C) 2003-2008 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.
24 extern int checksum_seed;
25 extern int append_mode;
27 int updating_basis_file;
29 static int false_alarms;
32 static int64 data_transfer;
34 static int total_false_alarms;
35 static int total_hash_hits;
36 static int total_matches;
38 extern struct stats stats;
40 #define TRADITIONAL_TABLESIZE (1<<16)
42 static uint32 tablesize;
43 static int32 *hash_table;
45 #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF)
46 #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16)
48 #define BIG_SUM2HASH(sum) ((sum)%tablesize)
50 static void build_hash_table(struct sum_struct *s)
52 static uint32 alloc_size;
55 /* Dynamically calculate the hash table size so that the hash load
56 * for big files is about 80%. A number greater than the traditional
57 * size must be odd or s2 will not be able to span the entire set. */
58 tablesize = (uint32)(s->count/8) * 10 + 11;
59 if (tablesize < TRADITIONAL_TABLESIZE)
60 tablesize = TRADITIONAL_TABLESIZE;
61 if (tablesize > alloc_size || tablesize < alloc_size - 16*1024) {
64 hash_table = new_array(int32, tablesize);
66 out_of_memory("build_hash_table");
67 alloc_size = tablesize;
70 memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]);
72 if (tablesize == TRADITIONAL_TABLESIZE) {
73 for (i = 0; i < s->count; i++) {
74 uint32 t = SUM2HASH(s->sums[i].sum1);
75 s->sums[i].chain = hash_table[t];
79 for (i = 0; i < s->count; i++) {
80 uint32 t = BIG_SUM2HASH(s->sums[i].sum1);
81 s->sums[i].chain = hash_table[t];
88 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 * @param i If >0, the number of a matched token. If 0, indicates we
101 * have only literal data.
103 static void matched(int f, struct sum_struct *s, struct map_struct *buf,
104 OFF_T offset, int32 i)
106 int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */
109 if (DEBUG_GTE(CHKSUM, 2) && i >= 0) {
111 "match at %.0f last_match=%.0f j=%d len=%ld n=%ld\n",
112 (double)offset, (double)last_match, i,
113 (long)s->sums[i].len, (long)n);
116 send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len);
120 stats.matched_data += s->sums[i].len;
124 for (j = 0; j < n; j += CHUNK_SIZE) {
125 int32 n1 = MIN(CHUNK_SIZE, n - j);
126 sum_update(map_ptr(buf, last_match + j, n1), n1);
130 last_match = offset + s->sums[i].len;
134 if (buf && INFO_GTE(PROGRESS, 1))
135 show_progress(last_match, buf->file_size);
139 static void hash_search(int f,struct sum_struct *s,
140 struct map_struct *buf, OFF_T len)
143 int32 k, want_i, backup;
144 char sum2[SUM_LENGTH];
149 /* want_i is used to encourage adjacent matches, allowing the RLL
150 * coding of the output to work more efficiently. */
153 if (DEBUG_GTE(CHKSUM, 2)) {
154 rprintf(FINFO, "hash search b=%ld len=%.0f\n",
155 (long)s->blength, (double)len);
158 k = (int32)MIN(len, (OFF_T)s->blength);
160 map = (schar *)map_ptr(buf, 0, k);
162 sum = get_checksum1((char *)map, k);
165 if (DEBUG_GTE(CHKSUM, 3))
166 rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
170 end = len + 1 - s->sums[s->count-1].len;
172 if (DEBUG_GTE(CHKSUM, 3)) {
173 rprintf(FINFO, "hash search s->blength=%ld len=%.0f count=%.0f\n",
174 (long)s->blength, (double)len, (double)s->count);
181 if (DEBUG_GTE(CHKSUM, 4)) {
182 rprintf(FINFO, "offset=%.0f sum=%04x%04x\n",
183 (double)offset, s2 & 0xFFFF, s1 & 0xFFFF);
186 if (tablesize == TRADITIONAL_TABLESIZE) {
187 if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0)
189 sum = (s1 & 0xffff) | (s2 << 16);
191 sum = (s1 & 0xffff) | (s2 << 16);
192 if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0)
200 if (sum != s->sums[i].sum1)
203 /* also make sure the two blocks are the same length */
204 l = (int32)MIN((OFF_T)s->blength, len-offset);
205 if (l != s->sums[i].len)
208 /* in-place: ensure chunk's offset is either >= our
209 * offset or that the data didn't move. */
210 if (updating_basis_file && s->sums[i].offset < offset
211 && !(s->sums[i].flags & SUMFLG_SAME_OFFSET))
214 if (DEBUG_GTE(CHKSUM, 3)) {
216 "potential match at %.0f i=%ld sum=%08x\n",
217 (double)offset, (long)i, sum);
221 map = (schar *)map_ptr(buf,offset,l);
222 get_checksum2((char *)map,l,sum2);
226 if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) {
231 /* When updating in-place, the best possible match is
232 * one with an identical offset, so we prefer that over
233 * the following want_i optimization. */
234 if (updating_basis_file) {
236 for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) {
237 if (s->sums[i2].offset != offset)
240 if (sum != s->sums[i2].sum1)
242 if (memcmp(sum2, s->sums[i2].sum2,
247 /* This chunk was at the same offset on
248 * both the sender and the receiver. */
249 s->sums[i].flags |= SUMFLG_SAME_OFFSET;
254 /* we've found a match, but now check to see
255 * if want_i can hint at a better match. */
256 if (i != want_i && want_i < s->count
257 && (!updating_basis_file || s->sums[want_i].offset >= offset
258 || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
259 && sum == s->sums[want_i].sum1
260 && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
261 /* we've found an adjacent match - the RLL coder
268 matched(f,s,buf,offset,i);
269 offset += s->sums[i].len - 1;
270 k = (int32)MIN((OFF_T)s->blength, len-offset);
271 map = (schar *)map_ptr(buf, offset, k);
272 sum = get_checksum1((char *)map, k);
277 } while ((i = s->sums[i].chain) >= 0);
280 backup = (int32)(offset - last_match);
281 /* We sometimes read 1 byte prior to last_match... */
285 /* Trim off the first byte from the checksum */
286 more = offset + k < len;
287 map = (schar *)map_ptr(buf, offset - backup, k + more + backup)
289 s1 -= map[0] + CHAR_OFFSET;
290 s2 -= k * (map[0]+CHAR_OFFSET);
292 /* Add on the next byte (if there is one) to the checksum */
294 s1 += map[k] + CHAR_OFFSET;
299 /* By matching early we avoid re-reading the
300 data 3 times in the case where a token
301 match comes a long way after last
302 match. The 3 reads are caused by the
303 running match, the checksum update and the
305 if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
306 matched(f, s, buf, offset - s->blength, -2);
307 } while (++offset < end);
309 matched(f, s, buf, len, -1);
310 map_ptr(buf, len-1, 1);
315 * Scan through a origin file, looking for sections that match
316 * checksums from the generator, and transmit either literal or token
319 * Also calculates the MD4 checksum of the whole file, using the md
320 * accumulator. This is transmitted with the file as protection
321 * against corruption on the wire.
323 * @param s Checksums received from the generator. If <tt>s->count ==
324 * 0</tt>, then there are actually no checksums for this file.
326 * @param len Length of the file to send.
328 void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
330 char file_sum[MAX_DIGEST_LEN];
339 sum_init(checksum_seed);
341 if (append_mode > 0) {
342 if (append_mode == 2) {
344 for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
345 if (buf && INFO_GTE(PROGRESS, 1))
346 show_progress(last_match, buf->file_size);
347 sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
351 if (last_match < s->flength) {
352 int32 n = (int32)(s->flength - last_match);
353 if (buf && INFO_GTE(PROGRESS, 1))
354 show_progress(last_match, buf->file_size);
355 sum_update(map_ptr(buf, last_match, n), n);
358 last_match = s->flength;
362 if (len > 0 && s->count > 0) {
365 if (DEBUG_GTE(CHKSUM, 2))
366 rprintf(FINFO,"built hash table\n");
368 hash_search(f, s, buf, len);
370 if (DEBUG_GTE(CHKSUM, 2))
371 rprintf(FINFO,"done hash search\n");
374 /* by doing this in pieces we avoid too many seeks */
375 for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE)
376 matched(f, s, buf, j, -2);
377 matched(f, s, buf, len, -1);
380 sum_len = sum_end(file_sum);
381 /* If we had a read error, send a bad checksum. */
382 if (buf && buf->status != 0)
385 if (DEBUG_GTE(CHKSUM, 2))
386 rprintf(FINFO,"sending file_sum\n");
387 write_buf(f, file_sum, sum_len);
389 if (DEBUG_GTE(CHKSUM, 2))
390 rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
391 false_alarms, hash_hits, matches);
393 total_hash_hits += hash_hits;
394 total_false_alarms += false_alarms;
395 total_matches += matches;
396 stats.literal_data += data_transfer;
399 void match_report(void)
401 if (!DEBUG_GTE(CHKSUM, 1))
405 "total: matches=%d hash_hits=%d false_alarms=%d data=%.0f\n",
406 total_matches, total_hash_hits, total_false_alarms,
407 (double)stats.literal_data);