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;
26 extern int checksum_len;
28 int updating_basis_file;
29 char sender_file_sum[MAX_DIGEST_LEN];
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;
94 * Transmit a literal and/or match token.
96 * This delightfully-named function is called either when we find a
97 * match and need to transmit all the unmatched data leading up to it,
98 * or when we get bored of accumulating literal data and just need to
99 * transmit it. As a result of this second case, it is called even if
100 * we have not matched at all!
102 * @param i If >0, the number of a matched token. If 0, indicates we
103 * have only literal data.
105 static void matched(int f, struct sum_struct *s, struct map_struct *buf,
106 OFF_T offset, int32 i)
108 int32 n = (int32)(offset - last_match); /* max value: block_size (int32) */
111 if (DEBUG_GTE(DELTASUM, 2) && i >= 0) {
113 "match at %s last_match=%s j=%d len=%ld n=%ld\n",
114 big_num(offset, 0), big_num(last_match, 0), i,
115 (long)s->sums[i].len, (long)n);
118 send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len);
122 stats.matched_data += s->sums[i].len;
126 for (j = 0; j < n; j += CHUNK_SIZE) {
127 int32 n1 = MIN(CHUNK_SIZE, n - j);
128 sum_update(map_ptr(buf, last_match + j, n1), n1);
132 last_match = offset + s->sums[i].len;
136 if (buf && INFO_GTE(PROGRESS, 1))
137 show_progress(last_match, buf->file_size);
141 static void hash_search(int f,struct sum_struct *s,
142 struct map_struct *buf, OFF_T len)
145 int32 k, want_i, backup;
146 char sum2[SUM_LENGTH];
151 /* want_i is used to encourage adjacent matches, allowing the RLL
152 * coding of the output to work more efficiently. */
155 if (DEBUG_GTE(DELTASUM, 2)) {
156 rprintf(FINFO, "hash search b=%ld len=%s\n",
157 (long)s->blength, big_num(len, 0));
160 k = (int32)MIN(len, (OFF_T)s->blength);
162 map = (schar *)map_ptr(buf, 0, k);
164 sum = get_checksum1((char *)map, k);
167 if (DEBUG_GTE(DELTASUM, 3))
168 rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
172 end = len + 1 - s->sums[s->count-1].len;
174 if (DEBUG_GTE(DELTASUM, 3)) {
175 rprintf(FINFO, "hash search s->blength=%ld len=%s count=%s\n",
176 (long)s->blength, big_num(len, 0), big_num(s->count, 0));
183 if (DEBUG_GTE(DELTASUM, 4)) {
184 rprintf(FINFO, "offset=%s sum=%04x%04x\n",
185 big_num(offset, 0), s2 & 0xFFFF, s1 & 0xFFFF);
188 if (tablesize == TRADITIONAL_TABLESIZE) {
189 if ((i = hash_table[SUM2HASH2(s1,s2)]) < 0)
191 sum = (s1 & 0xffff) | (s2 << 16);
193 sum = (s1 & 0xffff) | (s2 << 16);
194 if ((i = hash_table[BIG_SUM2HASH(sum)]) < 0)
202 if (sum != s->sums[i].sum1)
205 /* also make sure the two blocks are the same length */
206 l = (int32)MIN((OFF_T)s->blength, len-offset);
207 if (l != s->sums[i].len)
210 /* in-place: ensure chunk's offset is either >= our
211 * offset or that the data didn't move. */
212 if (updating_basis_file && s->sums[i].offset < offset
213 && !(s->sums[i].flags & SUMFLG_SAME_OFFSET))
216 if (DEBUG_GTE(DELTASUM, 3)) {
218 "potential match at %s i=%ld sum=%08x\n",
219 big_num(offset, 0), (long)i, sum);
223 map = (schar *)map_ptr(buf,offset,l);
224 get_checksum2((char *)map,l,sum2);
228 if (memcmp(sum2,s->sums[i].sum2,s->s2length) != 0) {
233 /* When updating in-place, the best possible match is
234 * one with an identical offset, so we prefer that over
235 * the following want_i optimization. */
236 if (updating_basis_file) {
238 for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) {
239 if (s->sums[i2].offset != offset)
242 if (sum != s->sums[i2].sum1)
244 if (memcmp(sum2, s->sums[i2].sum2,
249 /* This chunk was at the same offset on
250 * both the sender and the receiver. */
251 s->sums[i].flags |= SUMFLG_SAME_OFFSET;
256 /* we've found a match, but now check to see
257 * if want_i can hint at a better match. */
258 if (i != want_i && want_i < s->count
259 && (!updating_basis_file || s->sums[want_i].offset >= offset
260 || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
261 && sum == s->sums[want_i].sum1
262 && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
263 /* we've found an adjacent match - the RLL coder
270 matched(f,s,buf,offset,i);
271 offset += s->sums[i].len - 1;
272 k = (int32)MIN((OFF_T)s->blength, len-offset);
273 map = (schar *)map_ptr(buf, offset, k);
274 sum = get_checksum1((char *)map, k);
279 } while ((i = s->sums[i].chain) >= 0);
282 backup = (int32)(offset - last_match);
283 /* We sometimes read 1 byte prior to last_match... */
287 /* Trim off the first byte from the checksum */
288 more = offset + k < len;
289 map = (schar *)map_ptr(buf, offset - backup, k + more + backup)
291 s1 -= map[0] + CHAR_OFFSET;
292 s2 -= k * (map[0]+CHAR_OFFSET);
294 /* Add on the next byte (if there is one) to the checksum */
296 s1 += map[k] + CHAR_OFFSET;
301 /* By matching early we avoid re-reading the
302 data 3 times in the case where a token
303 match comes a long way after last
304 match. The 3 reads are caused by the
305 running match, the checksum update and the
307 if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
308 matched(f, s, buf, offset - s->blength, -2);
309 } while (++offset < end);
311 matched(f, s, buf, len, -1);
312 map_ptr(buf, len-1, 1);
317 * Scan through a origin file, looking for sections that match
318 * checksums from the generator, and transmit either literal or token
321 * Also calculates the MD4 checksum of the whole file, using the md
322 * accumulator. This is transmitted with the file as protection
323 * against corruption on the wire.
325 * @param s Checksums received from the generator. If <tt>s->count ==
326 * 0</tt>, then there are actually no checksums for this file.
328 * @param len Length of the file to send.
330 void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
338 sum_init(checksum_seed);
340 if (append_mode > 0) {
341 if (append_mode == 2) {
343 for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
344 if (buf && INFO_GTE(PROGRESS, 1))
345 show_progress(last_match, buf->file_size);
346 sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
350 if (last_match < s->flength) {
351 int32 n = (int32)(s->flength - last_match);
352 if (buf && INFO_GTE(PROGRESS, 1))
353 show_progress(last_match, buf->file_size);
354 sum_update(map_ptr(buf, last_match, n), n);
357 last_match = s->flength;
361 if (len > 0 && s->count > 0) {
364 if (DEBUG_GTE(DELTASUM, 2))
365 rprintf(FINFO,"built hash table\n");
367 hash_search(f, s, buf, len);
369 if (DEBUG_GTE(DELTASUM, 2))
370 rprintf(FINFO,"done hash search\n");
373 /* by doing this in pieces we avoid too many seeks */
374 for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE)
375 matched(f, s, buf, j, -2);
376 matched(f, s, buf, len, -1);
379 if (sum_end(sender_file_sum) != checksum_len)
380 overflow_exit("checksum_len"); /* Impossible... */
382 /* If we had a read error, send a bad checksum. We use all bits
383 * off as long as the checksum doesn't happen to be that, in
384 * which case we turn the last 0 bit into a 1. */
385 if (buf && buf->status != 0) {
387 for (i = 0; i < checksum_len && sender_file_sum[i] == 0; i++) {}
388 memset(sender_file_sum, 0, checksum_len);
389 if (i == checksum_len)
390 sender_file_sum[i-1]++;
393 if (DEBUG_GTE(DELTASUM, 2))
394 rprintf(FINFO,"sending file_sum\n");
395 write_buf(f, sender_file_sum, checksum_len);
397 if (DEBUG_GTE(DELTASUM, 2)) {
398 rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
399 false_alarms, hash_hits, matches);
402 total_hash_hits += hash_hits;
403 total_false_alarms += false_alarms;
404 total_matches += matches;
405 stats.literal_data += data_transfer;
408 void match_report(void)
410 if (!DEBUG_GTE(DELTASUM, 1))
414 "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n",
415 total_matches, total_hash_hits, total_false_alarms,
416 big_num(stats.literal_data, 0));