2 * Block matching used by the file-transfer code.
4 * Copyright (C) 1996 Andrew Tridgell
5 * Copyright (C) 1996 Paul Mackerras
6 * Copyright (C) 2003-2020 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 xfersum_type
;
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
;
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 * If i >= 0, the number of a matched token. If < 0, indicates we have
103 * only literal data. A -1 will send a 0-token-int too, and a -2 sends
104 * only literal data, w/o any token-int. */
105 static void matched(int f
, struct sum_struct
*s
, struct map_struct
*buf
, OFF_T offset
, int32 i
)
107 int32 n
= (int32
)(offset
- last_match
); /* max value: block_size (int32) */
110 if (DEBUG_GTE(DELTASUM
, 2) && i
>= 0) {
112 "match at %s last_match=%s j=%d len=%ld n=%ld\n",
113 big_num(offset
), big_num(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
&& INFO_GTE(PROGRESS
, 1))
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
, aligned_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. */
154 if (DEBUG_GTE(DELTASUM
, 2)) {
155 rprintf(FINFO
, "hash search b=%ld len=%s\n",
156 (long)s
->blength
, big_num(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
);
166 if (DEBUG_GTE(DELTASUM
, 3))
167 rprintf(FINFO
, "sum=%.8x k=%ld\n", sum
, (long)k
);
169 offset
= aligned_offset
= aligned_i
= 0;
171 end
= len
+ 1 - s
->sums
[s
->count
-1].len
;
173 if (DEBUG_GTE(DELTASUM
, 3)) {
174 rprintf(FINFO
, "hash search s->blength=%ld len=%s count=%s\n",
175 (long)s
->blength
, big_num(len
), big_num(s
->count
));
183 if (DEBUG_GTE(DELTASUM
, 4)) {
184 rprintf(FINFO
, "offset=%s sum=%04x%04x\n",
185 big_num(offset
), s2
& 0xFFFF, s1
& 0xFFFF);
188 if (tablesize
== TRADITIONAL_TABLESIZE
) {
189 hash_entry
= SUM2HASH2(s1
,s2
);
190 if ((i
= hash_table
[hash_entry
]) < 0)
192 sum
= (s1
& 0xffff) | (s2
<< 16);
194 sum
= (s1
& 0xffff) | (s2
<< 16);
195 hash_entry
= BIG_SUM2HASH(sum
);
196 if ((i
= hash_table
[hash_entry
]) < 0)
199 prev
= &hash_table
[hash_entry
];
205 /* When updating in-place, the chunk's offset must be
206 * either >= our offset or identical data at that offset.
207 * Remove any bypassed entries that we can never use. */
208 if (updating_basis_file
&& s
->sums
[i
].offset
< offset
209 && !(s
->sums
[i
].flags
& SUMFLG_SAME_OFFSET
)) {
210 *prev
= s
->sums
[i
].chain
;
213 prev
= &s
->sums
[i
].chain
;
215 if (sum
!= s
->sums
[i
].sum1
)
218 /* also make sure the two blocks are the same length */
219 l
= (int32
)MIN((OFF_T
)s
->blength
, len
-offset
);
220 if (l
!= s
->sums
[i
].len
)
223 if (DEBUG_GTE(DELTASUM
, 3)) {
225 "potential match at %s i=%ld sum=%08x\n",
226 big_num(offset
), (long)i
, sum
);
230 map
= (schar
*)map_ptr(buf
,offset
,l
);
231 get_checksum2((char *)map
,l
,sum2
);
235 if (memcmp(sum2
,s
->sums
[i
].sum2
,s
->s2length
) != 0) {
240 /* When updating in-place, the best possible match is
241 * one with an identical offset, so we prefer that over
242 * the adjacent want_i optimization. */
243 if (updating_basis_file
) {
244 /* All the generator's chunks start at blength boundaries. */
245 while (aligned_offset
< offset
) {
246 aligned_offset
+= s
->blength
;
249 if ((offset
== aligned_offset
250 || (sum
== 0 && l
== s
->blength
&& aligned_offset
+ l
<= len
))
251 && aligned_i
< s
->count
) {
252 if (i
!= aligned_i
) {
253 if (sum
!= s
->sums
[aligned_i
].sum1
254 || l
!= s
->sums
[aligned_i
].len
255 || memcmp(sum2
, s
->sums
[aligned_i
].sum2
, s
->s2length
) != 0)
259 if (offset
!= aligned_offset
) {
260 /* We've matched some zeros in a spot that is also zeros
261 * further along in the basis file, if we find zeros ahead
262 * in the sender's file, we'll output enough literal data
263 * to re-align with the basis file, and get back to seeking
264 * instead of writing. */
265 backup
= (int32
)(aligned_offset
- last_match
);
268 map
= (schar
*)map_ptr(buf
, aligned_offset
- backup
, l
+ backup
)
270 sum
= get_checksum1((char *)map
, l
);
271 if (sum
!= s
->sums
[i
].sum1
)
273 get_checksum2((char *)map
, l
, sum2
);
274 if (memcmp(sum2
, s
->sums
[i
].sum2
, s
->s2length
) != 0)
276 /* OK, we have a re-alignment match. Bump the offset
277 * forward to the new match point. */
278 offset
= aligned_offset
;
280 /* This identical chunk is in the same spot in the old and new file. */
281 s
->sums
[i
].flags
|= SUMFLG_SAME_OFFSET
;
287 /* we've found a match, but now check to see
288 * if want_i can hint at a better match. */
289 if (i
!= want_i
&& want_i
< s
->count
290 && (!updating_basis_file
|| s
->sums
[want_i
].offset
>= offset
291 || s
->sums
[want_i
].flags
& SUMFLG_SAME_OFFSET
)
292 && sum
== s
->sums
[want_i
].sum1
293 && memcmp(sum2
, s
->sums
[want_i
].sum2
, s
->s2length
) == 0) {
294 /* we've found an adjacent match - the RLL coder
300 matched(f
,s
,buf
,offset
,i
);
301 offset
+= s
->sums
[i
].len
- 1;
302 k
= (int32
)MIN((OFF_T
)s
->blength
, len
-offset
);
303 map
= (schar
*)map_ptr(buf
, offset
, k
);
304 sum
= get_checksum1((char *)map
, k
);
309 } while ((i
= s
->sums
[i
].chain
) >= 0);
312 backup
= (int32
)(offset
- last_match
);
313 /* We sometimes read 1 byte prior to last_match... */
317 /* Trim off the first byte from the checksum */
318 more
= offset
+ k
< len
;
319 map
= (schar
*)map_ptr(buf
, offset
- backup
, k
+ more
+ backup
) + backup
;
320 s1
-= map
[0] + CHAR_OFFSET
;
321 s2
-= k
* (map
[0]+CHAR_OFFSET
);
323 /* Add on the next byte (if there is one) to the checksum */
325 s1
+= map
[k
] + CHAR_OFFSET
;
330 /* By matching early we avoid re-reading the
331 data 3 times in the case where a token
332 match comes a long way after last
333 match. The 3 reads are caused by the
334 running match, the checksum update and the
336 if (backup
>= s
->blength
+CHUNK_SIZE
&& end
-offset
> CHUNK_SIZE
)
337 matched(f
, s
, buf
, offset
- s
->blength
, -2);
338 } while (++offset
< end
);
340 matched(f
, s
, buf
, len
, -1);
341 map_ptr(buf
, len
-1, 1);
346 * Scan through a origin file, looking for sections that match
347 * checksums from the generator, and transmit either literal or token
350 * Also calculates the MD4 checksum of the whole file, using the md
351 * accumulator. This is transmitted with the file as protection
352 * against corruption on the wire.
354 * @param s Checksums received from the generator. If <tt>s->count ==
355 * 0</tt>, then there are actually no checksums for this file.
357 * @param len Length of the file to send.
359 void match_sums(int f
, struct sum_struct
*s
, struct map_struct
*buf
, OFF_T len
)
369 sum_init(xfersum_type
, checksum_seed
);
371 if (append_mode
> 0) {
372 if (append_mode
== 2) {
374 for (j
= CHUNK_SIZE
; j
< s
->flength
; j
+= CHUNK_SIZE
) {
375 if (buf
&& INFO_GTE(PROGRESS
, 1))
376 show_progress(last_match
, buf
->file_size
);
377 sum_update(map_ptr(buf
, last_match
, CHUNK_SIZE
),
381 if (last_match
< s
->flength
) {
382 int32 n
= (int32
)(s
->flength
- last_match
);
383 if (buf
&& INFO_GTE(PROGRESS
, 1))
384 show_progress(last_match
, buf
->file_size
);
385 sum_update(map_ptr(buf
, last_match
, n
), n
);
388 last_match
= s
->flength
;
392 if (len
> 0 && s
->count
> 0) {
395 if (DEBUG_GTE(DELTASUM
, 2))
396 rprintf(FINFO
,"built hash table\n");
398 hash_search(f
, s
, buf
, len
);
400 if (DEBUG_GTE(DELTASUM
, 2))
401 rprintf(FINFO
,"done hash search\n");
404 /* by doing this in pieces we avoid too many seeks */
405 for (j
= last_match
+ CHUNK_SIZE
; j
< len
; j
+= CHUNK_SIZE
)
406 matched(f
, s
, buf
, j
, -2);
407 matched(f
, s
, buf
, len
, -1);
410 sum_len
= sum_end(sender_file_sum
);
412 /* If we had a read error, send a bad checksum. We use all bits
413 * off as long as the checksum doesn't happen to be that, in
414 * which case we turn the last 0 bit into a 1. */
415 if (buf
&& buf
->status
!= 0) {
417 for (i
= 0; i
< sum_len
&& sender_file_sum
[i
] == 0; i
++) {}
418 memset(sender_file_sum
, 0, sum_len
);
420 sender_file_sum
[i
-1]++;
423 if (DEBUG_GTE(DELTASUM
, 2))
424 rprintf(FINFO
,"sending file_sum\n");
425 write_buf(f
, sender_file_sum
, sum_len
);
427 if (DEBUG_GTE(DELTASUM
, 2)) {
428 rprintf(FINFO
, "false_alarms=%d hash_hits=%d matches=%d\n",
429 false_alarms
, hash_hits
, matches
);
432 total_hash_hits
+= hash_hits
;
433 total_false_alarms
+= false_alarms
;
434 total_matches
+= matches
;
435 stats
.literal_data
+= data_transfer
;
438 void match_report(void)
440 if (!DEBUG_GTE(DELTASUM
, 1))
444 "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n",
445 total_matches
, total_hash_hits
, total_false_alarms
,
446 big_num(stats
.literal_data
));