2 * Block matching used by the file-transfer code.
4 * Copyright (C) 1996 Andrew Tridgell
5 * Copyright (C) 1996 Paul Mackerras
6 * Copyright (C) 2003-2014 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
;
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
,
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
), big_num(last_match
), 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
)
144 OFF_T offset
, aligned_offset
, end
;
145 int32 k
, want_i
, aligned_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
));
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
);
170 offset
= aligned_offset
= aligned_i
= 0;
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
), 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 hash_entry
= SUM2HASH2(s1
,s2
);
191 if ((i
= hash_table
[hash_entry
]) < 0)
193 sum
= (s1
& 0xffff) | (s2
<< 16);
195 sum
= (s1
& 0xffff) | (s2
<< 16);
196 hash_entry
= BIG_SUM2HASH(sum
);
197 if ((i
= hash_table
[hash_entry
]) < 0)
200 prev
= &hash_table
[hash_entry
];
206 /* When updating in-place, the chunk's offset must be
207 * either >= our offset or identical data at that offset.
208 * Remove any bypassed entries that we can never use. */
209 if (updating_basis_file
&& s
->sums
[i
].offset
< offset
210 && !(s
->sums
[i
].flags
& SUMFLG_SAME_OFFSET
)) {
211 *prev
= s
->sums
[i
].chain
;
214 prev
= &s
->sums
[i
].chain
;
216 if (sum
!= s
->sums
[i
].sum1
)
219 /* also make sure the two blocks are the same length */
220 l
= (int32
)MIN((OFF_T
)s
->blength
, len
-offset
);
221 if (l
!= s
->sums
[i
].len
)
224 if (DEBUG_GTE(DELTASUM
, 3)) {
226 "potential match at %s i=%ld sum=%08x\n",
227 big_num(offset
), (long)i
, sum
);
231 map
= (schar
*)map_ptr(buf
,offset
,l
);
232 get_checksum2((char *)map
,l
,sum2
);
236 if (memcmp(sum2
,s
->sums
[i
].sum2
,s
->s2length
) != 0) {
241 /* When updating in-place, the best possible match is
242 * one with an identical offset, so we prefer that over
243 * the adjacent want_i optimization. */
244 if (updating_basis_file
) {
245 /* All the generator's chunks start at blength boundaries. */
246 while (aligned_offset
< offset
) {
247 aligned_offset
+= s
->blength
;
250 if ((offset
== aligned_offset
251 || (sum
== 0 && l
== s
->blength
&& aligned_offset
+ l
<= len
))
252 && aligned_i
< s
->count
) {
253 if (i
!= aligned_i
) {
254 if (sum
!= s
->sums
[aligned_i
].sum1
255 || l
!= s
->sums
[aligned_i
].len
256 || memcmp(sum2
, s
->sums
[aligned_i
].sum2
, s
->s2length
) != 0)
260 if (offset
!= aligned_offset
) {
261 /* We've matched some zeros in a spot that is also zeros
262 * further along in the basis file, if we find zeros ahead
263 * in the sender's file, we'll output enough literal data
264 * to re-align with the basis file, and get back to seeking
265 * instead of writing. */
266 backup
= (int32
)(aligned_offset
- last_match
);
269 map
= (schar
*)map_ptr(buf
, aligned_offset
- backup
, l
+ backup
)
271 sum
= get_checksum1((char *)map
, l
);
272 if (sum
!= s
->sums
[i
].sum1
)
274 get_checksum2((char *)map
, l
, sum2
);
275 if (memcmp(sum2
, s
->sums
[i
].sum2
, s
->s2length
) != 0)
277 /* OK, we have a re-alignment match. Bump the offset
278 * forward to the new match point. */
279 offset
= aligned_offset
;
281 /* This identical chunk is in the same spot in the old and new file. */
282 s
->sums
[i
].flags
|= SUMFLG_SAME_OFFSET
;
288 /* we've found a match, but now check to see
289 * if want_i can hint at a better match. */
290 if (i
!= want_i
&& want_i
< s
->count
291 && (!updating_basis_file
|| s
->sums
[want_i
].offset
>= offset
292 || s
->sums
[want_i
].flags
& SUMFLG_SAME_OFFSET
)
293 && sum
== s
->sums
[want_i
].sum1
294 && memcmp(sum2
, s
->sums
[want_i
].sum2
, s
->s2length
) == 0) {
295 /* we've found an adjacent match - the RLL coder
301 matched(f
,s
,buf
,offset
,i
);
302 offset
+= s
->sums
[i
].len
- 1;
303 k
= (int32
)MIN((OFF_T
)s
->blength
, len
-offset
);
304 map
= (schar
*)map_ptr(buf
, offset
, k
);
305 sum
= get_checksum1((char *)map
, k
);
310 } while ((i
= s
->sums
[i
].chain
) >= 0);
313 backup
= (int32
)(offset
- last_match
);
314 /* We sometimes read 1 byte prior to last_match... */
318 /* Trim off the first byte from the checksum */
319 more
= offset
+ k
< len
;
320 map
= (schar
*)map_ptr(buf
, offset
- backup
, k
+ more
+ backup
)
322 s1
-= map
[0] + CHAR_OFFSET
;
323 s2
-= k
* (map
[0]+CHAR_OFFSET
);
325 /* Add on the next byte (if there is one) to the checksum */
327 s1
+= map
[k
] + CHAR_OFFSET
;
332 /* By matching early we avoid re-reading the
333 data 3 times in the case where a token
334 match comes a long way after last
335 match. The 3 reads are caused by the
336 running match, the checksum update and the
338 if (backup
>= s
->blength
+CHUNK_SIZE
&& end
-offset
> CHUNK_SIZE
)
339 matched(f
, s
, buf
, offset
- s
->blength
, -2);
340 } while (++offset
< end
);
342 matched(f
, s
, buf
, len
, -1);
343 map_ptr(buf
, len
-1, 1);
348 * Scan through a origin file, looking for sections that match
349 * checksums from the generator, and transmit either literal or token
352 * Also calculates the MD4 checksum of the whole file, using the md
353 * accumulator. This is transmitted with the file as protection
354 * against corruption on the wire.
356 * @param s Checksums received from the generator. If <tt>s->count ==
357 * 0</tt>, then there are actually no checksums for this file.
359 * @param len Length of the file to send.
361 void match_sums(int f
, struct sum_struct
*s
, struct map_struct
*buf
, OFF_T len
)
369 sum_init(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 if (sum_end(sender_file_sum
) != checksum_len
)
411 overflow_exit("checksum_len"); /* Impossible... */
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
< checksum_len
&& sender_file_sum
[i
] == 0; i
++) {}
419 memset(sender_file_sum
, 0, checksum_len
);
420 if (i
== checksum_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
, checksum_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
));