2 * Block matching used by the file-transfer code.
4 * Copyright (C) 1996 Andrew Tridgell
5 * Copyright (C) 1996 Paul Mackerras
6 * Copyright (C) 2003-2023 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
;
28 extern struct name_num_item
*xfer_sum_nni
;
29 extern int xfer_sum_len
;
31 int updating_basis_file
;
32 char sender_file_sum
[MAX_DIGEST_LEN
];
34 static int false_alarms
;
37 static int64 data_transfer
;
39 static int total_false_alarms
;
40 static int total_hash_hits
;
41 static int total_matches
;
43 extern struct stats stats
;
45 #define TRADITIONAL_TABLESIZE (1<<16)
47 static uint32 tablesize
;
48 static int32
*hash_table
;
50 #define SUM2HASH2(s1,s2) (((s1) + (s2)) & 0xFFFF)
51 #define SUM2HASH(sum) SUM2HASH2((sum)&0xFFFF,(sum)>>16)
53 #define BIG_SUM2HASH(sum) ((sum)%tablesize)
55 static void build_hash_table(struct sum_struct
*s
)
57 static uint32 alloc_size
;
60 /* Dynamically calculate the hash table size so that the hash load
61 * for big files is about 80%. A number greater than the traditional
62 * size must be odd or s2 will not be able to span the entire set. */
63 tablesize
= (uint32
)(s
->count
/8) * 10 + 11;
64 if (tablesize
< TRADITIONAL_TABLESIZE
)
65 tablesize
= TRADITIONAL_TABLESIZE
;
66 if (tablesize
> alloc_size
|| tablesize
< alloc_size
- 16*1024) {
69 hash_table
= new_array(int32
, tablesize
);
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
[MAX_DIGEST_LEN
];
150 // prevent possible memory leaks
151 memset(sum2
, 0, sizeof sum2
);
153 /* want_i is used to encourage adjacent matches, allowing the RLL
154 * coding of the output to work more efficiently. */
157 if (DEBUG_GTE(DELTASUM
, 2)) {
158 rprintf(FINFO
, "hash search b=%ld len=%s\n",
159 (long)s
->blength
, big_num(len
));
162 k
= (int32
)MIN(len
, (OFF_T
)s
->blength
);
164 map
= (schar
*)map_ptr(buf
, 0, k
);
166 sum
= get_checksum1((char *)map
, k
);
169 if (DEBUG_GTE(DELTASUM
, 3))
170 rprintf(FINFO
, "sum=%.8x k=%ld\n", sum
, (long)k
);
172 offset
= aligned_offset
= aligned_i
= 0;
174 end
= len
+ 1 - s
->sums
[s
->count
-1].len
;
176 if (DEBUG_GTE(DELTASUM
, 3)) {
177 rprintf(FINFO
, "hash search s->blength=%ld len=%s count=%s\n",
178 (long)s
->blength
, big_num(len
), big_num(s
->count
));
186 if (DEBUG_GTE(DELTASUM
, 4)) {
187 rprintf(FINFO
, "offset=%s sum=%04x%04x\n",
188 big_num(offset
), s2
& 0xFFFF, s1
& 0xFFFF);
191 if (tablesize
== TRADITIONAL_TABLESIZE
) {
192 hash_entry
= SUM2HASH2(s1
,s2
);
193 if ((i
= hash_table
[hash_entry
]) < 0)
195 sum
= (s1
& 0xffff) | (s2
<< 16);
197 sum
= (s1
& 0xffff) | (s2
<< 16);
198 hash_entry
= BIG_SUM2HASH(sum
);
199 if ((i
= hash_table
[hash_entry
]) < 0)
202 prev
= &hash_table
[hash_entry
];
208 /* When updating in-place, the chunk's offset must be
209 * either >= our offset or identical data at that offset.
210 * Remove any bypassed entries that we can never use. */
211 if (updating_basis_file
&& s
->sums
[i
].offset
< offset
212 && !(s
->sums
[i
].flags
& SUMFLG_SAME_OFFSET
)) {
213 *prev
= s
->sums
[i
].chain
;
216 prev
= &s
->sums
[i
].chain
;
218 if (sum
!= s
->sums
[i
].sum1
)
221 /* also make sure the two blocks are the same length */
222 l
= (int32
)MIN((OFF_T
)s
->blength
, len
-offset
);
223 if (l
!= s
->sums
[i
].len
)
226 if (DEBUG_GTE(DELTASUM
, 3)) {
228 "potential match at %s i=%ld sum=%08x\n",
229 big_num(offset
), (long)i
, sum
);
233 map
= (schar
*)map_ptr(buf
,offset
,l
);
234 get_checksum2((char *)map
,l
,sum2
);
238 if (memcmp(sum2
, sum2_at(s
, i
), s
->s2length
) != 0) {
243 /* When updating in-place, the best possible match is
244 * one with an identical offset, so we prefer that over
245 * the adjacent want_i optimization. */
246 if (updating_basis_file
) {
247 /* All the generator's chunks start at blength boundaries. */
248 while (aligned_offset
< offset
) {
249 aligned_offset
+= s
->blength
;
252 if ((offset
== aligned_offset
253 || (sum
== 0 && l
== s
->blength
&& aligned_offset
+ l
<= len
))
254 && aligned_i
< s
->count
) {
255 if (i
!= aligned_i
) {
256 if (sum
!= s
->sums
[aligned_i
].sum1
257 || l
!= s
->sums
[aligned_i
].len
258 || memcmp(sum2
, sum2_at(s
, aligned_i
), s
->s2length
) != 0)
262 if (offset
!= aligned_offset
) {
263 /* We've matched some zeros in a spot that is also zeros
264 * further along in the basis file, if we find zeros ahead
265 * in the sender's file, we'll output enough literal data
266 * to re-align with the basis file, and get back to seeking
267 * instead of writing. */
268 backup
= (int32
)(aligned_offset
- last_match
);
271 map
= (schar
*)map_ptr(buf
, aligned_offset
- backup
, l
+ backup
)
273 sum
= get_checksum1((char *)map
, l
);
274 if (sum
!= s
->sums
[i
].sum1
)
276 get_checksum2((char *)map
, l
, sum2
);
277 if (memcmp(sum2
, sum2_at(s
, i
), s
->s2length
) != 0)
279 /* OK, we have a re-alignment match. Bump the offset
280 * forward to the new match point. */
281 offset
= aligned_offset
;
283 /* This identical chunk is in the same spot in the old and new file. */
284 s
->sums
[i
].flags
|= SUMFLG_SAME_OFFSET
;
290 /* we've found a match, but now check to see
291 * if want_i can hint at a better match. */
292 if (i
!= want_i
&& want_i
< s
->count
293 && (!updating_basis_file
|| s
->sums
[want_i
].offset
>= offset
294 || s
->sums
[want_i
].flags
& SUMFLG_SAME_OFFSET
)
295 && sum
== s
->sums
[want_i
].sum1
296 && memcmp(sum2
, sum2_at(s
, want_i
), s
->s2length
) == 0) {
297 /* we've found an adjacent match - the RLL coder
303 matched(f
,s
,buf
,offset
,i
);
304 offset
+= s
->sums
[i
].len
- 1;
305 k
= (int32
)MIN((OFF_T
)s
->blength
, len
-offset
);
306 map
= (schar
*)map_ptr(buf
, offset
, k
);
307 sum
= get_checksum1((char *)map
, k
);
312 } while ((i
= s
->sums
[i
].chain
) >= 0);
315 backup
= (int32
)(offset
- last_match
);
316 /* We sometimes read 1 byte prior to last_match... */
320 /* Trim off the first byte from the checksum */
321 more
= offset
+ k
< len
;
322 map
= (schar
*)map_ptr(buf
, offset
- backup
, k
+ more
+ backup
) + backup
;
323 s1
-= map
[0] + CHAR_OFFSET
;
324 s2
-= k
* (map
[0]+CHAR_OFFSET
);
326 /* Add on the next byte (if there is one) to the checksum */
328 s1
+= map
[k
] + CHAR_OFFSET
;
333 /* By matching early we avoid re-reading the
334 data 3 times in the case where a token
335 match comes a long way after last
336 match. The 3 reads are caused by the
337 running match, the checksum update and the
339 if (backup
>= s
->blength
+CHUNK_SIZE
&& end
-offset
> CHUNK_SIZE
)
340 matched(f
, s
, buf
, offset
- s
->blength
, -2);
341 } while (++offset
< end
);
343 matched(f
, s
, buf
, len
, -1);
344 map_ptr(buf
, len
-1, 1);
349 * Scan through a origin file, looking for sections that match
350 * checksums from the generator, and transmit either literal or token
353 * Also calculates the MD4 checksum of the whole file, using the md
354 * accumulator. This is transmitted with the file as protection
355 * against corruption on the wire.
357 * @param s Checksums received from the generator. If <tt>s->count ==
358 * 0</tt>, then there are actually no checksums for this file.
360 * @param len Length of the file to send.
362 void match_sums(int f
, struct sum_struct
*s
, struct map_struct
*buf
, OFF_T len
)
370 sum_init(xfer_sum_nni
, checksum_seed
);
372 if (append_mode
> 0) {
373 if (append_mode
== 2) {
375 for (j
= CHUNK_SIZE
; j
< s
->flength
; j
+= CHUNK_SIZE
) {
376 if (buf
&& INFO_GTE(PROGRESS
, 1))
377 show_progress(last_match
, buf
->file_size
);
378 sum_update(map_ptr(buf
, last_match
, CHUNK_SIZE
),
382 if (last_match
< s
->flength
) {
383 int32 n
= (int32
)(s
->flength
- last_match
);
384 if (buf
&& INFO_GTE(PROGRESS
, 1))
385 show_progress(last_match
, buf
->file_size
);
386 sum_update(map_ptr(buf
, last_match
, n
), n
);
389 last_match
= s
->flength
;
393 if (len
> 0 && s
->count
> 0) {
396 if (DEBUG_GTE(DELTASUM
, 2))
397 rprintf(FINFO
,"built hash table\n");
399 hash_search(f
, s
, buf
, len
);
401 if (DEBUG_GTE(DELTASUM
, 2))
402 rprintf(FINFO
,"done hash search\n");
405 /* by doing this in pieces we avoid too many seeks */
406 for (j
= last_match
+ CHUNK_SIZE
; j
< len
; j
+= CHUNK_SIZE
)
407 matched(f
, s
, buf
, j
, -2);
408 matched(f
, s
, buf
, len
, -1);
411 sum_end(sender_file_sum
);
413 /* If we had a read error, send a bad checksum. We use all bits
414 * off as long as the checksum doesn't happen to be that, in
415 * which case we turn the last 0 bit into a 1. */
416 if (buf
&& buf
->status
!= 0) {
418 for (i
= 0; i
< xfer_sum_len
&& sender_file_sum
[i
] == 0; i
++) {}
419 memset(sender_file_sum
, 0, xfer_sum_len
);
420 if (i
== xfer_sum_len
)
421 sender_file_sum
[i
-1]++;
424 if (DEBUG_GTE(DELTASUM
, 2))
425 rprintf(FINFO
,"sending file_sum\n");
426 write_buf(f
, sender_file_sum
, xfer_sum_len
);
428 if (DEBUG_GTE(DELTASUM
, 2)) {
429 rprintf(FINFO
, "false_alarms=%d hash_hits=%d matches=%d\n",
430 false_alarms
, hash_hits
, matches
);
433 total_hash_hits
+= hash_hits
;
434 total_false_alarms
+= false_alarms
;
435 total_matches
+= matches
;
436 stats
.literal_data
+= data_transfer
;
439 void match_report(void)
441 if (!DEBUG_GTE(DELTASUM
, 1))
445 "total: matches=%d hash_hits=%d false_alarms=%d data=%s\n",
446 total_matches
, total_hash_hits
, total_false_alarms
,
447 big_num(stats
.literal_data
));