2 * Block matching used by the file-transfer code.
4 * Copyright (C) 1996 Andrew Tridgell
5 * Copyright (C) 1996 Paul Mackerras
6 * Copyright (C) 2003-2007 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 TABLESIZE (1<<16)
44 static int32 *hash_table;
46 #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF)
47 #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16)
49 static void build_hash_table(struct sum_struct *s)
54 hash_table = new_array(int32, TABLESIZE);
56 out_of_memory("build_hash_table");
59 memset(hash_table, 0xFF, TABLESIZE * sizeof hash_table[0]);
61 for (i = 0; i < s->count; i++) {
62 uint32 t = SUM2HASH(s->sums[i].sum1);
63 s->sums[i].chain = hash_table[t];
69 static OFF_T last_match;
73 * Transmit a literal and/or match token.
75 * This delightfully-named function is called either when we find a
76 * match and need to transmit all the unmatched data leading up to it,
77 * or when we get bored of accumulating literal data and just need to
78 * transmit it. As a result of this second case, it is called even if
79 * we have not matched at all!
81 * @param i If >0, the number of a matched token. If 0, indicates we
82 * have only literal data.
84 static void matched(int f, struct sum_struct *s, struct map_struct *buf,
85 OFF_T offset, int32 i)
87 int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */
90 if (verbose > 2 && i >= 0) {
92 "match at %.0f last_match=%.0f j=%d len=%ld n=%ld\n",
93 (double)offset, (double)last_match, i,
94 (long)s->sums[i].len, (long)n);
97 send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len);
101 stats.matched_data += s->sums[i].len;
105 for (j = 0; j < n; j += CHUNK_SIZE) {
106 int32 n1 = MIN(CHUNK_SIZE, n - j);
107 sum_update(map_ptr(buf, last_match + j, n1), n1);
111 last_match = offset + s->sums[i].len;
115 if (buf && do_progress)
116 show_progress(last_match, buf->file_size);
120 static void hash_search(int f,struct sum_struct *s,
121 struct map_struct *buf, OFF_T len)
124 int32 k, want_i, backup;
125 char sum2[SUM_LENGTH];
130 /* want_i is used to encourage adjacent matches, allowing the RLL
131 * coding of the output to work more efficiently. */
135 rprintf(FINFO, "hash search b=%ld len=%.0f\n",
136 (long)s->blength, (double)len);
139 k = (int32)MIN(len, (OFF_T)s->blength);
141 map = (schar *)map_ptr(buf, 0, k);
143 sum = get_checksum1((char *)map, k);
147 rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
151 end = len + 1 - s->sums[s->count-1].len;
154 rprintf(FINFO, "hash search s->blength=%ld len=%.0f count=%.0f\n",
155 (long)s->blength, (double)len, (double)s->count);
163 rprintf(FINFO, "offset=%.0f sum=%04x%04x\n",
164 (double)offset, s2 & 0xFFFF, s1 & 0xFFFF);
167 i = hash_table[SUM2HASH2(s1,s2)];
171 sum = (s1 & 0xffff) | (s2 << 16);
176 if (sum != s->sums[i].sum1)
179 /* also make sure the two blocks are the same length */
180 l = (int32)MIN((OFF_T)s->blength, len-offset);
181 if (l != s->sums[i].len)
184 /* in-place: ensure chunk's offset is either >= our
185 * offset or that the data didn't move. */
186 if (updating_basis_file && s->sums[i].offset < offset
187 && !(s->sums[i].flags & SUMFLG_SAME_OFFSET))
192 "potential match at %.0f i=%ld sum=%08x\n",
193 (double)offset, (long)i, sum);
197 map = (schar *)map_ptr(buf,offset,l);
198 get_checksum2((char *)map,l,sum2);
202 if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) {
207 /* When updating in-place, the best possible match is
208 * one with an identical offset, so we prefer that over
209 * the following want_i optimization. */
210 if (updating_basis_file) {
212 for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) {
213 if (s->sums[i2].offset != offset)
216 if (sum != s->sums[i2].sum1)
218 if (memcmp(sum2, s->sums[i2].sum2,
223 /* This chunk was at the same offset on
224 * both the sender and the receiver. */
225 s->sums[i].flags |= SUMFLG_SAME_OFFSET;
230 /* we've found a match, but now check to see
231 * if want_i can hint at a better match. */
232 if (i != want_i && want_i < s->count
233 && (!updating_basis_file || s->sums[want_i].offset >= offset
234 || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
235 && sum == s->sums[want_i].sum1
236 && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
237 /* we've found an adjacent match - the RLL coder
244 matched(f,s,buf,offset,i);
245 offset += s->sums[i].len - 1;
246 k = (int32)MIN((OFF_T)s->blength, len-offset);
247 map = (schar *)map_ptr(buf, offset, k);
248 sum = get_checksum1((char *)map, k);
253 } while ((i = s->sums[i].chain) >= 0);
256 backup = (int32)(offset - last_match);
257 /* We sometimes read 1 byte prior to last_match... */
261 /* Trim off the first byte from the checksum */
262 more = offset + k < len;
263 map = (schar *)map_ptr(buf, offset - backup, k + more + backup)
265 s1 -= map[0] + CHAR_OFFSET;
266 s2 -= k * (map[0]+CHAR_OFFSET);
268 /* Add on the next byte (if there is one) to the checksum */
270 s1 += map[k] + CHAR_OFFSET;
275 /* By matching early we avoid re-reading the
276 data 3 times in the case where a token
277 match comes a long way after last
278 match. The 3 reads are caused by the
279 running match, the checksum update and the
281 if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
282 matched(f, s, buf, offset - s->blength, -2);
283 } while (++offset < end);
285 matched(f, s, buf, len, -1);
286 map_ptr(buf, len-1, 1);
291 * Scan through a origin file, looking for sections that match
292 * checksums from the generator, and transmit either literal or token
295 * Also calculates the MD4 checksum of the whole file, using the md
296 * accumulator. This is transmitted with the file as protection
297 * against corruption on the wire.
299 * @param s Checksums received from the generator. If <tt>s->count ==
300 * 0</tt>, then there are actually no checksums for this file.
302 * @param len Length of the file to send.
304 void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
306 char file_sum[MAX_DIGEST_LEN];
315 sum_init(checksum_seed);
317 if (append_mode > 0) {
318 if (append_mode == 2) {
320 for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
321 if (buf && do_progress)
322 show_progress(last_match, buf->file_size);
323 sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
327 if (last_match < s->flength) {
328 int32 n = (int32)(s->flength - last_match);
329 if (buf && do_progress)
330 show_progress(last_match, buf->file_size);
331 sum_update(map_ptr(buf, last_match, n), n);
334 last_match = s->flength;
338 if (len > 0 && s->count > 0) {
342 rprintf(FINFO,"built hash table\n");
344 hash_search(f, s, buf, len);
347 rprintf(FINFO,"done hash search\n");
350 /* by doing this in pieces we avoid too many seeks */
351 for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE)
352 matched(f, s, buf, j, -2);
353 matched(f, s, buf, len, -1);
356 sum_len = sum_end(file_sum);
357 /* If we had a read error, send a bad checksum. */
358 if (buf && buf->status != 0)
362 rprintf(FINFO,"sending file_sum\n");
363 write_buf(f, file_sum, sum_len);
366 rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
367 false_alarms, hash_hits, matches);
369 total_hash_hits += hash_hits;
370 total_false_alarms += false_alarms;
371 total_matches += matches;
372 stats.literal_data += data_transfer;
375 void match_report(void)
381 "total: matches=%d hash_hits=%d false_alarms=%d data=%.0f\n",
382 total_matches, total_hash_hits, total_false_alarms,
383 (double)stats.literal_data);