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 checksum_seed;
26 extern int append_mode;
27 extern int checksum_len;
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);
69 out_of_memory("build_hash_table");
70 alloc_size = tablesize;
73 memset(hash_table, 0xFF, tablesize * sizeof hash_table[0]);
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];
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];
91 static OFF_T last_match;
95 * Transmit a literal and/or match token.
97 * This delightfully-named function is called either when we find a
98 * match and need to transmit all the unmatched data leading up to it,
99 * or when we get bored of accumulating literal data and just need to
100 * transmit it. As a result of this second case, it is called even if
101 * we have not matched at all!
103 * @param i If >0, the number of a matched token. If 0, indicates we
104 * have only literal data.
106 static void matched(int f, struct sum_struct *s, struct map_struct *buf,
107 OFF_T offset, int32 i)
109 int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */
112 if (DEBUG_GTE(DELTASUM, 2) && i >= 0) {
114 "match at %s last_match=%s j=%d len=%ld n=%ld\n",
115 big_num(offset), big_num(last_match), i,
116 (long)s->sums[i].len, (long)n);
119 send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len);
123 stats.matched_data += s->sums[i].len;
127 for (j = 0; j < n; j += CHUNK_SIZE) {
128 int32 n1 = MIN(CHUNK_SIZE, n - j);
129 sum_update(map_ptr(buf, last_match + j, n1), n1);
133 last_match = offset + s->sums[i].len;
137 if (buf && INFO_GTE(PROGRESS, 1))
138 show_progress(last_match, buf->file_size);
142 static void hash_search(int f,struct sum_struct *s,
143 struct map_struct *buf, OFF_T len)
146 int32 k, want_i, backup;
147 char sum2[SUM_LENGTH];
152 /* want_i is used to encourage adjacent matches, allowing the RLL
153 * coding of the output to work more efficiently. */
156 if (DEBUG_GTE(DELTASUM, 2)) {
157 rprintf(FINFO, "hash search b=%ld len=%s\n",
158 (long)s->blength, big_num(len));
161 k = (int32)MIN(len, (OFF_T)s->blength);
163 map = (schar *)map_ptr(buf, 0, k);
165 sum = get_checksum1((char *)map, k);
168 if (DEBUG_GTE(DELTASUM, 3))
169 rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
173 end = len + 1 - s->sums[s->count-1].len;
175 if (DEBUG_GTE(DELTASUM, 3)) {
176 rprintf(FINFO, "hash search s->blength=%ld len=%s count=%s\n",
177 (long)s->blength, big_num(len), big_num(s->count));
184 if (DEBUG_GTE(DELTASUM, 4)) {
185 rprintf(FINFO, "offset=%s sum=%04x%04x\n",
186 big_num(offset), s2 & 0xFFFF, s1 & 0xFFFF);
189 if (tablesize == TRADITIONAL_TABLESIZE) {
190 if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0)
192 sum = (s1 & 0xffff) | (s2 << 16);
194 sum = (s1 & 0xffff) | (s2 << 16);
195 if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0)
203 if (sum != s->sums[i].sum1)
206 /* also make sure the two blocks are the same length */
207 l = (int32)MIN((OFF_T)s->blength, len-offset);
208 if (l != s->sums[i].len)
211 /* in-place: ensure chunk's offset is either >= our
212 * offset or that the data didn't move. */
213 if (updating_basis_file && s->sums[i].offset < offset
214 && !(s->sums[i].flags & SUMFLG_SAME_OFFSET))
217 if (DEBUG_GTE(DELTASUM, 3)) {
219 "potential match at %s i=%ld sum=%08x\n",
220 big_num(offset), (long)i, sum);
224 map = (schar *)map_ptr(buf,offset,l);
225 get_checksum2((char *)map,l,sum2);
229 if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) {
234 /* When updating in-place, the best possible match is
235 * one with an identical offset, so we prefer that over
236 * the following want_i optimization. */
237 if (updating_basis_file) {
239 for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) {
240 if (s->sums[i2].offset != offset)
243 if (sum != s->sums[i2].sum1)
245 if (memcmp(sum2, s->sums[i2].sum2,
250 /* This chunk was at the same offset on
251 * both the sender and the receiver. */
252 s->sums[i].flags |= SUMFLG_SAME_OFFSET;
257 /* we've found a match, but now check to see
258 * if want_i can hint at a better match. */
259 if (i != want_i && want_i < s->count
260 && (!updating_basis_file || s->sums[want_i].offset >= offset
261 || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
262 && sum == s->sums[want_i].sum1
263 && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
264 /* we've found an adjacent match - the RLL coder
271 matched(f,s,buf,offset,i);
272 offset += s->sums[i].len - 1;
273 k = (int32)MIN((OFF_T)s->blength, len-offset);
274 map = (schar *)map_ptr(buf, offset, k);
275 sum = get_checksum1((char *)map, k);
280 } while ((i = s->sums[i].chain) >= 0);
283 backup = (int32)(offset - last_match);
284 /* We sometimes read 1 byte prior to last_match... */
288 /* Trim off the first byte from the checksum */
289 more = offset + k < len;
290 map = (schar *)map_ptr(buf, offset - backup, k + more + backup)
292 s1 -= map[0] + CHAR_OFFSET;
293 s2 -= k * (map[0]+CHAR_OFFSET);
295 /* Add on the next byte (if there is one) to the checksum */
297 s1 += map[k] + CHAR_OFFSET;
302 /* By matching early we avoid re-reading the
303 data 3 times in the case where a token
304 match comes a long way after last
305 match. The 3 reads are caused by the
306 running match, the checksum update and the
308 if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
309 matched(f, s, buf, offset - s->blength, -2);
310 } while (++offset < end);
312 matched(f, s, buf, len, -1);
313 map_ptr(buf, len-1, 1);
318 * Scan through a origin file, looking for sections that match
319 * checksums from the generator, and transmit either literal or token
322 * Also calculates the MD4 checksum of the whole file, using the md
323 * accumulator. This is transmitted with the file as protection
324 * against corruption on the wire.
326 * @param s Checksums received from the generator. If <tt>s->count ==
327 * 0</tt>, then there are actually no checksums for this file.
329 * @param len Length of the file to send.
331 void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T 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(DELTASUM, 2))
366 rprintf(FINFO,"built hash table\n");
368 hash_search(f, s, buf, len);
370 if (DEBUG_GTE(DELTASUM, 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 if (sum_end(sender_file_sum) != checksum_len)
381 overflow_exit("checksum_len"); /* Impossible... */
383 /* If we had a read error, send a bad checksum. We use all bits
384 * off as long as the checksum doesn't happen to be that, in
385 * which case we turn the last 0 bit into a 1. */
386 if (buf && buf->status != 0) {
388 for (i = 0; i < checksum_len && sender_file_sum[i] == 0; i++) {}
389 memset(sender_file_sum, 0, checksum_len);
390 if (i == checksum_len)
391 sender_file_sum[i-1]++;
394 if (DEBUG_GTE(DELTASUM, 2))
395 rprintf(FINFO,"sending file_sum\n");
396 write_buf(f, sender_file_sum, checksum_len);
398 if (DEBUG_GTE(DELTASUM, 2)) {
399 rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
400 false_alarms, hash_hits, matches);
403 total_hash_hits += hash_hits;
404 total_false_alarms += false_alarms;
405 total_matches += matches;
406 stats.literal_data += data_transfer;
409 void match_report(void)
411 if (!DEBUG_GTE(DELTASUM, 1))
415 "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n",
416 total_matches, total_hash_hits, total_false_alarms,
417 big_num(stats.literal_data));