1 /* vi: set sw=4 ts=4: */
2 /* Small bzip2 deflate implementation, by Rob Landley (rob@landley.net).
4 Based on bzip2 decompression code by Julian R Seward (jseward@acm.org),
5 which also acknowledges contributions by Mike Burrows, David Wheeler,
6 Peter Fenwick, Alistair Moffat, Radford Neal, Ian H. Witten,
7 Robert Sedgewick, and Jon L. Bentley.
9 Licensed under GPLv2 or later, see file LICENSE in this source tree.
13 Size and speed optimizations by Manuel Novoa III (mjn3@codepoet.org).
15 More efficient reading of Huffman codes, a streamlined read_bunzip()
16 function, and various other tweaks. In (limited) tests, approximately
17 20% faster than bzcat on x86 and about 10% faster on arm.
19 Note that about 2/3 of the time is spent in read_bunzip() reversing
20 the Burrows-Wheeler transformation. Much of that time is delay
21 resulting from cache misses.
23 (2010 update by vda: profiled "bzcat <84mbyte.bz2 >/dev/null"
24 on x86-64 CPU with L2 > 1M: get_next_block is hotter than read_bunzip:
25 %time seconds calls function
26 71.01 12.69 444 get_next_block
27 28.65 5.12 93065 read_bunzip
28 00.22 0.04 7736490 get_bits
29 00.11 0.02 47 dealloc_bunzip
30 00.00 0.00 93018 full_write
34 I would ask that anyone benefiting from this work, especially those
35 using it in commercial products, consider making a donation to my local
36 non-profit hospice organization (www.hospiceacadiana.com) in the name of
37 the woman I loved, Toni W. Hagan, who passed away Feb. 12, 2003.
43 #include "bb_archive.h"
45 /* Constants for Huffman coding */
47 #define GROUP_SIZE 50 /* 64 would have been more efficient */
48 #define MAX_HUFCODE_BITS 20 /* Longest Huffman code allowed */
49 #define MAX_SYMBOLS 258 /* 256 literals + RUNA + RUNB */
53 /* Status return values */
55 #define RETVAL_LAST_BLOCK (-1)
56 #define RETVAL_NOT_BZIP_DATA (-2)
57 #define RETVAL_UNEXPECTED_INPUT_EOF (-3)
58 #define RETVAL_SHORT_WRITE (-4)
59 #define RETVAL_DATA_ERROR (-5)
60 #define RETVAL_OUT_OF_MEMORY (-6)
61 #define RETVAL_OBSOLETE_INPUT (-7)
63 /* Other housekeeping constants */
64 #define IOBUF_SIZE 4096
66 /* This is what we know about each Huffman coding group */
68 /* We have an extra slot at the end of limit[] for a sentinel value. */
69 int limit[MAX_HUFCODE_BITS+1], base[MAX_HUFCODE_BITS], permute[MAX_SYMBOLS];
73 /* Structure holding all the housekeeping data, including IO buffers and
74 * memory that persists between calls to bunzip
75 * Found the most used member:
76 * cat this_file.c | sed -e 's/"/ /g' -e "s/'/ /g" | xargs -n1 \
77 * | grep 'bd->' | sed 's/^.*bd->/bd->/' | sort | $PAGER
78 * and moved it (inbufBitCount) to offset 0.
81 /* I/O tracking data (file handles, buffers, positions, etc.) */
82 unsigned inbufBitCount, inbufBits;
83 int in_fd, out_fd, inbufCount, inbufPos /*, outbufPos*/;
84 uint8_t *inbuf /*,*outbuf*/;
86 /* State for interrupting output loop */
87 int writeCopies, writePos, writeRunCountdown, writeCount;
88 int writeCurrent; /* actually a uint8_t */
90 /* The CRC values stored in the block header and calculated from the data */
91 uint32_t headerCRC, totalCRC, writeCRC;
93 /* Intermediate buffer and its size (in bytes) */
97 /* For I/O error handling */
100 /* Big things go last (register-relative addressing can be larger for big offsets) */
101 uint32_t crc32Table[256];
102 uint8_t selectors[32768]; /* nSelectors=15 bits */
103 struct group_data groups[MAX_GROUPS]; /* Huffman coding tables */
105 /* typedef struct bunzip_data bunzip_data; -- done in .h file */
108 /* Return the next nnn bits of input. All reads from the compressed input
109 are done through this function. All reads are big endian */
110 static unsigned get_bits(bunzip_data *bd, int bits_wanted)
113 /* Cache bd->inbufBitCount in a CPU register (hopefully): */
114 int bit_count = bd->inbufBitCount;
116 /* If we need to get more data from the byte buffer, do so. (Loop getting
117 one byte at a time to enforce endianness and avoid unaligned access.) */
118 while (bit_count < bits_wanted) {
120 /* If we need to read more data from file into byte buffer, do so */
121 if (bd->inbufPos == bd->inbufCount) {
122 /* if "no input fd" case: in_fd == -1, read fails, we jump */
123 bd->inbufCount = read(bd->in_fd, bd->inbuf, IOBUF_SIZE);
124 if (bd->inbufCount <= 0)
125 longjmp(bd->jmpbuf, RETVAL_UNEXPECTED_INPUT_EOF);
129 /* Avoid 32-bit overflow (dump bit buffer to top of output) */
130 if (bit_count >= 24) {
131 bits = bd->inbufBits & ((1 << bit_count) - 1);
132 bits_wanted -= bit_count;
133 bits <<= bits_wanted;
137 /* Grab next 8 bits of input from buffer. */
138 bd->inbufBits = (bd->inbufBits << 8) | bd->inbuf[bd->inbufPos++];
142 /* Calculate result */
143 bit_count -= bits_wanted;
144 bd->inbufBitCount = bit_count;
145 bits |= (bd->inbufBits >> bit_count) & ((1 << bits_wanted) - 1);
150 /* Unpacks the next block and sets up for the inverse Burrows-Wheeler step. */
151 static int get_next_block(bunzip_data *bd)
153 struct group_data *hufGroup;
154 int dbufCount, dbufSize, groupCount, *base, *limit, selector,
155 i, j, t, runPos, symCount, symTotal, nSelectors, byteCount[256];
156 int runCnt = runCnt; /* for compiler */
157 uint8_t uc, symToByte[256], mtfSymbol[256], *selectors;
162 dbufSize = bd->dbufSize;
163 selectors = bd->selectors;
165 /* In bbox, we are ok with aborting through setjmp which is set up in start_bunzip */
167 /* Reset longjmp I/O error handling */
168 i = setjmp(bd->jmpbuf);
172 /* Read in header signature and CRC, then validate signature.
173 (last block signature means CRC is for whole file, return now) */
174 i = get_bits(bd, 24);
175 j = get_bits(bd, 24);
176 bd->headerCRC = get_bits(bd, 32);
177 if ((i == 0x177245) && (j == 0x385090)) return RETVAL_LAST_BLOCK;
178 if ((i != 0x314159) || (j != 0x265359)) return RETVAL_NOT_BZIP_DATA;
180 /* We can add support for blockRandomised if anybody complains. There was
181 some code for this in busybox 1.0.0-pre3, but nobody ever noticed that
182 it didn't actually work. */
183 if (get_bits(bd, 1)) return RETVAL_OBSOLETE_INPUT;
184 origPtr = get_bits(bd, 24);
185 if ((int)origPtr > dbufSize) return RETVAL_DATA_ERROR;
187 /* mapping table: if some byte values are never used (encoding things
188 like ascii text), the compression code removes the gaps to have fewer
189 symbols to deal with, and writes a sparse bitfield indicating which
190 values were present. We make a translation table to convert the symbols
191 back to the corresponding bytes. */
194 t = get_bits(bd, 16);
197 unsigned inner_map = get_bits(bd, 16);
199 if (inner_map & (1 << 15))
200 symToByte[symTotal++] = i;
210 /* How many different Huffman coding groups does this block use? */
211 groupCount = get_bits(bd, 3);
212 if (groupCount < 2 || groupCount > MAX_GROUPS)
213 return RETVAL_DATA_ERROR;
215 /* nSelectors: Every GROUP_SIZE many symbols we select a new Huffman coding
216 group. Read in the group selector list, which is stored as MTF encoded
217 bit runs. (MTF=Move To Front, as each value is used it's moved to the
218 start of the list.) */
219 for (i = 0; i < groupCount; i++)
221 nSelectors = get_bits(bd, 15);
223 return RETVAL_DATA_ERROR;
224 for (i = 0; i < nSelectors; i++) {
228 while (get_bits(bd, 1)) {
229 if (n >= groupCount) return RETVAL_DATA_ERROR;
232 /* Decode MTF to get the next selector */
233 tmp_byte = mtfSymbol[n];
235 mtfSymbol[n + 1] = mtfSymbol[n];
236 mtfSymbol[0] = selectors[i] = tmp_byte;
239 /* Read the Huffman coding tables for each group, which code for symTotal
240 literal symbols, plus two run symbols (RUNA, RUNB) */
241 symCount = symTotal + 2;
242 for (j = 0; j < groupCount; j++) {
243 uint8_t length[MAX_SYMBOLS];
244 /* 8 bits is ALMOST enough for temp[], see below */
245 unsigned temp[MAX_HUFCODE_BITS+1];
246 int minLen, maxLen, pp, len_m1;
248 /* Read Huffman code lengths for each symbol. They're stored in
249 a way similar to mtf; record a starting value for the first symbol,
250 and an offset from the previous value for every symbol after that.
251 (Subtracting 1 before the loop and then adding it back at the end is
252 an optimization that makes the test inside the loop simpler: symbol
253 length 0 becomes negative, so an unsigned inequality catches it.) */
254 len_m1 = get_bits(bd, 5) - 1;
255 for (i = 0; i < symCount; i++) {
258 if ((unsigned)len_m1 > (MAX_HUFCODE_BITS-1))
259 return RETVAL_DATA_ERROR;
261 /* If first bit is 0, stop. Else second bit indicates whether
262 to increment or decrement the value. Optimization: grab 2
263 bits and unget the second if the first was 0. */
264 two_bits = get_bits(bd, 2);
270 /* Add one if second bit 1, else subtract 1. Avoids if/else */
271 len_m1 += (((two_bits+1) & 2) - 1);
274 /* Correct for the initial -1, to get the final symbol length */
275 length[i] = len_m1 + 1;
278 /* Find largest and smallest lengths in this group */
279 minLen = maxLen = length[0];
280 for (i = 1; i < symCount; i++) {
281 if (length[i] > maxLen) maxLen = length[i];
282 else if (length[i] < minLen) minLen = length[i];
285 /* Calculate permute[], base[], and limit[] tables from length[].
287 * permute[] is the lookup table for converting Huffman coded symbols
288 * into decoded symbols. base[] is the amount to subtract from the
289 * value of a Huffman symbol of a given length when using permute[].
291 * limit[] indicates the largest numerical value a symbol with a given
292 * number of bits can have. This is how the Huffman codes can vary in
293 * length: each code with a value>limit[length] needs another bit.
295 hufGroup = bd->groups + j;
296 hufGroup->minLen = minLen;
297 hufGroup->maxLen = maxLen;
299 /* Note that minLen can't be smaller than 1, so we adjust the base
300 and limit array pointers so we're not always wasting the first
301 entry. We do this again when using them (during symbol decoding). */
302 base = hufGroup->base - 1;
303 limit = hufGroup->limit - 1;
305 /* Calculate permute[]. Concurently, initialize temp[] and limit[]. */
307 for (i = minLen; i <= maxLen; i++) {
309 temp[i] = limit[i] = 0;
310 for (k = 0; k < symCount; k++)
312 hufGroup->permute[pp++] = k;
315 /* Count symbols coded for at each bit length */
316 /* NB: in pathological cases, temp[8] can end ip being 256.
317 * That's why uint8_t is too small for temp[]. */
318 for (i = 0; i < symCount; i++) temp[length[i]]++;
320 /* Calculate limit[] (the largest symbol-coding value at each bit
321 * length, which is (previous limit<<1)+symbols at this level), and
322 * base[] (number of symbols to ignore at each bit length, which is
323 * limit minus the cumulative count of symbols coded for already). */
325 for (i = minLen; i < maxLen;) {
326 unsigned temp_i = temp[i];
330 /* We read the largest possible symbol size and then unget bits
331 after determining how many we need, and those extra bits could
332 be set to anything. (They're noise from future symbols.) At
333 each level we're really only interested in the first few bits,
334 so here we set all the trailing to-be-ignored bits to 1 so they
335 don't affect the value>limit[length] comparison. */
336 limit[i] = (pp << (maxLen - i)) - 1;
341 limit[maxLen] = pp + temp[maxLen] - 1;
342 limit[maxLen+1] = INT_MAX; /* Sentinel value for reading next sym. */
346 /* We've finished reading and digesting the block header. Now read this
347 block's Huffman coded symbols from the file and undo the Huffman coding
348 and run length encoding, saving the result into dbuf[dbufCount++] = uc */
350 /* Initialize symbol occurrence counters and symbol Move To Front table */
351 /*memset(byteCount, 0, sizeof(byteCount)); - smaller, but slower */
352 for (i = 0; i < 256; i++) {
354 mtfSymbol[i] = (uint8_t)i;
357 /* Loop through compressed symbols. */
359 runPos = dbufCount = selector = 0;
363 /* Fetch next Huffman coding group from list. */
364 symCount = GROUP_SIZE - 1;
365 if (selector >= nSelectors) return RETVAL_DATA_ERROR;
366 hufGroup = bd->groups + selectors[selector++];
367 base = hufGroup->base - 1;
368 limit = hufGroup->limit - 1;
371 /* Read next Huffman-coded symbol. */
373 /* Note: It is far cheaper to read maxLen bits and back up than it is
374 to read minLen bits and then add additional bit at a time, testing
375 as we go. Because there is a trailing last block (with file CRC),
376 there is no danger of the overread causing an unexpected EOF for a
377 valid compressed file.
380 /* As a further optimization, we do the read inline
381 (falling back to a call to get_bits if the buffer runs dry).
384 while ((new_cnt = bd->inbufBitCount - hufGroup->maxLen) < 0) {
385 /* bd->inbufBitCount < hufGroup->maxLen */
386 if (bd->inbufPos == bd->inbufCount) {
387 nextSym = get_bits(bd, hufGroup->maxLen);
390 bd->inbufBits = (bd->inbufBits << 8) | bd->inbuf[bd->inbufPos++];
391 bd->inbufBitCount += 8;
393 bd->inbufBitCount = new_cnt; /* "bd->inbufBitCount -= hufGroup->maxLen;" */
394 nextSym = (bd->inbufBits >> new_cnt) & ((1 << hufGroup->maxLen) - 1);
396 } else { /* unoptimized equivalent */
397 nextSym = get_bits(bd, hufGroup->maxLen);
399 /* Figure how many bits are in next symbol and unget extras */
400 i = hufGroup->minLen;
401 while (nextSym > limit[i]) ++i;
402 j = hufGroup->maxLen - i;
404 return RETVAL_DATA_ERROR;
405 bd->inbufBitCount += j;
407 /* Huffman decode value to get nextSym (with bounds checking) */
408 nextSym = (nextSym >> j) - base[i];
409 if ((unsigned)nextSym >= MAX_SYMBOLS)
410 return RETVAL_DATA_ERROR;
411 nextSym = hufGroup->permute[nextSym];
413 /* We have now decoded the symbol, which indicates either a new literal
414 byte, or a repeated run of the most recent literal byte. First,
415 check if nextSym indicates a repeated run, and if so loop collecting
416 how many times to repeat the last literal. */
417 if ((unsigned)nextSym <= SYMBOL_RUNB) { /* RUNA or RUNB */
419 /* If this is the start of a new run, zero out counter */
425 /* Neat trick that saves 1 symbol: instead of or-ing 0 or 1 at
426 each bit position, add 1 or 2 instead. For example,
427 1011 is 1<<0 + 1<<1 + 2<<2. 1010 is 2<<0 + 2<<1 + 1<<2.
428 You can make any bit pattern that way using 1 less symbol than
429 the basic or 0/1 method (except all bits 0, which would use no
430 symbols, but a run of length 0 doesn't mean anything in this
431 context). Thus space is saved. */
432 runCnt += (runPos << nextSym); /* +runPos if RUNA; +2*runPos if RUNB */
433 if (runPos < dbufSize) runPos <<= 1;
434 goto end_of_huffman_loop;
437 /* When we hit the first non-run symbol after a run, we now know
438 how many times to repeat the last literal, so append that many
439 copies to our buffer of decoded symbols (dbuf) now. (The last
440 literal used is the one at the head of the mtfSymbol array.) */
443 if (dbufCount + runCnt >= dbufSize) return RETVAL_DATA_ERROR;
444 tmp_byte = symToByte[mtfSymbol[0]];
445 byteCount[tmp_byte] += runCnt;
446 while (--runCnt >= 0) dbuf[dbufCount++] = (uint32_t)tmp_byte;
450 /* Is this the terminating symbol? */
451 if (nextSym > symTotal) break;
453 /* At this point, nextSym indicates a new literal character. Subtract
454 one to get the position in the MTF array at which this literal is
455 currently to be found. (Note that the result can't be -1 or 0,
456 because 0 and 1 are RUNA and RUNB. But another instance of the
457 first symbol in the mtf array, position 0, would have been handled
458 as part of a run above. Therefore 1 unused mtf position minus
459 2 non-literal nextSym values equals -1.) */
460 if (dbufCount >= dbufSize) return RETVAL_DATA_ERROR;
464 /* Adjust the MTF array. Since we typically expect to move only a
465 * small number of symbols, and are bound by 256 in any case, using
466 * memmove here would typically be bigger and slower due to function
467 * call overhead and other assorted setup costs. */
469 mtfSymbol[i] = mtfSymbol[i-1];
474 /* We have our literal byte. Save it into dbuf. */
476 dbuf[dbufCount++] = (uint32_t)uc;
478 /* Skip group initialization if we're not done with this group. Done
479 * this way to avoid compiler warning. */
481 if (--symCount >= 0) goto continue_this_group;
484 /* At this point, we've read all the Huffman-coded symbols (and repeated
485 runs) for this block from the input stream, and decoded them into the
486 intermediate buffer. There are dbufCount many decoded bytes in dbuf[].
487 Now undo the Burrows-Wheeler transform on dbuf.
488 See http://dogma.net/markn/articles/bwt/bwt.htm
491 /* Turn byteCount into cumulative occurrence counts of 0 to n-1. */
493 for (i = 0; i < 256; i++) {
494 int tmp_count = j + byteCount[i];
499 /* Figure out what order dbuf would be in if we sorted it. */
500 for (i = 0; i < dbufCount; i++) {
501 uint8_t tmp_byte = (uint8_t)dbuf[i];
502 int tmp_count = byteCount[tmp_byte];
503 dbuf[tmp_count] |= (i << 8);
504 byteCount[tmp_byte] = tmp_count + 1;
507 /* Decode first byte by hand to initialize "previous" byte. Note that it
508 doesn't get output, and if the first three characters are identical
509 it doesn't qualify as a run (hence writeRunCountdown=5). */
512 if ((int)origPtr >= dbufCount) return RETVAL_DATA_ERROR;
514 bd->writeCurrent = (uint8_t)tmp;
515 bd->writePos = (tmp >> 8);
516 bd->writeRunCountdown = 5;
518 bd->writeCount = dbufCount;
523 /* Undo Burrows-Wheeler transform on intermediate buffer to produce output.
524 If start_bunzip was initialized with out_fd=-1, then up to len bytes of
525 data are written to outbuf. Return value is number of bytes written or
526 error (all errors are negative numbers). If out_fd!=-1, outbuf and len
527 are ignored, data is written to out_fd and return is RETVAL_OK or error.
529 NB: read_bunzip returns < 0 on error, or the number of *unfilled* bytes
530 in outbuf. IOW: on EOF returns len ("all bytes are not filled"), not 0.
531 (Why? This allows to get rid of one local variable)
533 int FAST_FUNC read_bunzip(bunzip_data *bd, char *outbuf, int len)
535 const uint32_t *dbuf;
536 int pos, current, previous;
539 /* If we already have error/end indicator, return it */
540 if (bd->writeCount < 0)
541 return bd->writeCount;
545 /* Register-cached state (hopefully): */
547 current = bd->writeCurrent;
548 CRC = bd->writeCRC; /* small loss on x86-32 (not enough regs), win on x86-64 */
550 /* We will always have pending decoded data to write into the output
551 buffer unless this is the very first call (in which case we haven't
552 Huffman-decoded a block into the intermediate buffer yet). */
553 if (bd->writeCopies) {
556 /* Inside the loop, writeCopies means extra copies (beyond 1) */
559 /* Loop outputting bytes */
562 /* If the output buffer is full, save cached state and return */
565 * Use of "goto" instead of keeping code here
566 * helps compiler to realize this. */
570 /* Write next byte into output buffer, updating CRC */
572 CRC = (CRC << 8) ^ bd->crc32Table[(CRC >> 24) ^ current];
574 /* Loop now if we're outputting multiple copies of this byte */
575 if (bd->writeCopies) {
576 /* Unlikely branch */
577 /*--bd->writeCopies;*/
579 /* Same, but (ab)using other existing --writeCopies operation
580 * (and this if() compiles into just test+branch pair): */
581 goto dec_writeCopies;
584 if (--bd->writeCount < 0)
585 break; /* input block is fully consumed, need next one */
587 /* Follow sequence vector to undo Burrows-Wheeler transform */
590 current = (uint8_t)pos;
593 /* After 3 consecutive copies of the same byte, the 4th
594 * is a repeat count. We count down from 4 instead
595 * of counting up because testing for non-zero is faster */
596 if (--bd->writeRunCountdown != 0) {
597 if (current != previous)
598 bd->writeRunCountdown = 4;
600 /* Unlikely branch */
601 /* We have a repeated run, this byte indicates the count */
602 bd->writeCopies = current;
604 bd->writeRunCountdown = 5;
606 /* Sometimes there are just 3 bytes (run length 0) */
607 if (!bd->writeCopies) goto decode_next_byte;
609 /* Subtract the 1 copy we'd output anyway to get extras */
614 /* Decompression of this input block completed successfully */
615 bd->writeCRC = CRC = ~CRC;
616 bd->totalCRC = ((bd->totalCRC << 1) | (bd->totalCRC >> 31)) ^ CRC;
618 /* If this block had a CRC error, force file level CRC error */
619 if (CRC != bd->headerCRC) {
620 bd->totalCRC = bd->headerCRC + 1;
621 return RETVAL_LAST_BLOCK;
625 /* Refill the intermediate buffer by Huffman-decoding next block of input */
627 int r = get_next_block(bd);
628 if (r) { /* error/end */
630 return (r != RETVAL_LAST_BLOCK) ? r : len;
636 current = bd->writeCurrent;
637 goto decode_next_byte;
640 /* Output buffer is full, save cached state and return */
642 bd->writeCurrent = current;
650 /* Allocate the structure, read file header. If in_fd==-1, inbuf must contain
651 a complete bunzip file (len bytes long). If in_fd!=-1, inbuf and len are
652 ignored, and data is read from file handle into temporary buffer. */
654 /* Because bunzip2 is used for help text unpacking, and because bb_show_usage()
655 should work for NOFORK applets too, we must be extremely careful to not leak
657 int FAST_FUNC start_bunzip(bunzip_data **bdp, int in_fd,
658 const void *inbuf, int len)
663 BZh0 = ('B' << 24) + ('Z' << 16) + ('h' << 8) + '0',
664 h0 = ('h' << 8) + '0',
667 /* Figure out how much data to allocate */
668 i = sizeof(bunzip_data);
669 if (in_fd != -1) i += IOBUF_SIZE;
671 /* Allocate bunzip_data. Most fields initialize to zero. */
672 bd = *bdp = xzalloc(i);
674 /* Setup input buffer */
677 /* in this case, bd->inbuf is read-only */
678 bd->inbuf = (void*)inbuf; /* cast away const-ness */
680 bd->inbuf = (uint8_t*)(bd + 1);
681 memcpy(bd->inbuf, inbuf, len);
683 bd->inbufCount = len;
685 /* Init the CRC32 table (big endian) */
686 crc32_filltable(bd->crc32Table, 1);
688 /* Setup for I/O error handling via longjmp */
689 i = setjmp(bd->jmpbuf);
692 /* Ensure that file starts with "BZh['1'-'9']." */
693 /* Update: now caller verifies 1st two bytes, makes .gz/.bz2
694 * integration easier */
696 /* i = get_bits(bd, 32); */
697 /* if ((unsigned)(i - BZh0 - 1) >= 9) return RETVAL_NOT_BZIP_DATA; */
698 i = get_bits(bd, 16);
699 if ((unsigned)(i - h0 - 1) >= 9) return RETVAL_NOT_BZIP_DATA;
701 /* Fourth byte (ascii '1'-'9') indicates block size in units of 100k of
702 uncompressed data. Allocate intermediate buffer for block. */
703 /* bd->dbufSize = 100000 * (i - BZh0); */
704 bd->dbufSize = 100000 * (i - h0);
706 /* Cannot use xmalloc - may leak bd in NOFORK case! */
707 bd->dbuf = malloc_or_warn(bd->dbufSize * sizeof(bd->dbuf[0]));
715 void FAST_FUNC dealloc_bunzip(bunzip_data *bd)
722 /* Decompress src_fd to dst_fd. Stops at end of bzip data, not end of file. */
723 IF_DESKTOP(long long) int FAST_FUNC
724 unpack_bz2_stream(transformer_aux_data_t *aux, int src_fd, int dst_fd)
726 IF_DESKTOP(long long total_written = 0;)
732 if (check_signature16(aux, src_fd, BZIP2_MAGIC))
735 outbuf = xmalloc(IOBUF_SIZE);
737 while (1) { /* "Process one BZ... stream" loop */
739 i = start_bunzip(&bd, src_fd, outbuf + 2, len);
742 while (1) { /* "Produce some output bytes" loop */
743 i = read_bunzip(bd, outbuf, IOBUF_SIZE);
744 if (i < 0) /* error? */
746 i = IOBUF_SIZE - i; /* number of bytes produced */
747 if (i == 0) /* EOF? */
749 if (i != full_write(dst_fd, outbuf, i)) {
750 bb_error_msg("short write");
751 i = RETVAL_SHORT_WRITE;
754 IF_DESKTOP(total_written += i;)
758 if (i != RETVAL_LAST_BLOCK
759 /* Observed case when i == RETVAL_OK:
760 * "bzcat z.bz2", where "z.bz2" is a bzipped zero-length file
761 * (to be exact, z.bz2 is exactly these 14 bytes:
762 * 42 5a 68 39 17 72 45 38 50 90 00 00 00 00).
766 bb_error_msg("bunzip error %d", i);
769 if (bd->headerCRC != bd->totalCRC) {
770 bb_error_msg("CRC error");
774 /* Successfully unpacked one BZ stream */
777 /* Do we have "BZ..." after last processed byte?
778 * pbzip2 (parallelized bzip2) produces such files.
780 len = bd->inbufCount - bd->inbufPos;
781 memcpy(outbuf, &bd->inbuf[bd->inbufPos], len);
783 if (safe_read(src_fd, outbuf + len, 2 - len) != 2 - len)
787 if (*(uint16_t*)outbuf != BZIP2_MAGIC) /* "BZ"? */
797 return i ? i : IF_DESKTOP(total_written) + 0;
802 static char *const bunzip_errors[] = {
803 NULL, "Bad file checksum", "Not bzip data",
804 "Unexpected input EOF", "Unexpected output EOF", "Data error",
805 "Out of memory", "Obsolete (pre 0.9.5) bzip format not supported"
808 /* Dumb little test thing, decompress stdin to stdout */
809 int main(int argc, char **argv)
814 int i = unpack_bz2_stream(0, 1);
816 fprintf(stderr, "%s\n", bunzip_errors[-i]);
817 else if (read(STDIN_FILENO, &c, 1))
818 fprintf(stderr, "Trailing garbage ignored\n");