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 tarball for details.
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_unzip() reversing
20 the Burrows-Wheeler transformation. Much of that time is delay
21 resulting from cache misses.
23 I would ask that anyone benefiting from this work, especially those
24 using it in commercial products, consider making a donation to my local
25 non-profit hospice organization (www.hospiceacadiana.com) in the name of
26 the woman I loved, Toni W. Hagan, who passed away Feb. 12, 2003.
32 #include "unarchive.h"
34 /* Constants for Huffman coding */
36 #define GROUP_SIZE 50 /* 64 would have been more efficient */
37 #define MAX_HUFCODE_BITS 20 /* Longest Huffman code allowed */
38 #define MAX_SYMBOLS 258 /* 256 literals + RUNA + RUNB */
42 /* Status return values */
44 #define RETVAL_LAST_BLOCK (-1)
45 #define RETVAL_NOT_BZIP_DATA (-2)
46 #define RETVAL_UNEXPECTED_INPUT_EOF (-3)
47 #define RETVAL_SHORT_WRITE (-4)
48 #define RETVAL_DATA_ERROR (-5)
49 #define RETVAL_OUT_OF_MEMORY (-6)
50 #define RETVAL_OBSOLETE_INPUT (-7)
52 /* Other housekeeping constants */
53 #define IOBUF_SIZE 4096
55 /* This is what we know about each Huffman coding group */
57 /* We have an extra slot at the end of limit[] for a sentinel value. */
58 int limit[MAX_HUFCODE_BITS+1], base[MAX_HUFCODE_BITS], permute[MAX_SYMBOLS];
62 /* Structure holding all the housekeeping data, including IO buffers and
63 * memory that persists between calls to bunzip
64 * Found the most used member:
65 * cat this_file.c | sed -e 's/"/ /g' -e "s/'/ /g" | xargs -n1 \
66 * | grep 'bd->' | sed 's/^.*bd->/bd->/' | sort | $PAGER
67 * and moved it (inbufBitCount) to offset 0.
71 /* I/O tracking data (file handles, buffers, positions, etc.) */
72 unsigned inbufBitCount, inbufBits;
73 int in_fd, out_fd, inbufCount, inbufPos /*, outbufPos*/;
74 unsigned char *inbuf /*,*outbuf*/;
76 /* State for interrupting output loop */
77 int writeCopies, writePos, writeRunCountdown, writeCount, writeCurrent;
79 /* The CRC values stored in the block header and calculated from the data */
80 uint32_t headerCRC, totalCRC, writeCRC;
82 /* Intermediate buffer and its size (in bytes) */
83 unsigned *dbuf, dbufSize;
85 /* For I/O error handling */
88 /* Big things go last (register-relative addressing can be larger for big offsets) */
89 uint32_t crc32Table[256];
90 unsigned char selectors[32768]; /* nSelectors=15 bits */
91 struct group_data groups[MAX_GROUPS]; /* Huffman coding tables */
93 /* typedef struct bunzip_data bunzip_data; -- done in .h file */
96 /* Return the next nnn bits of input. All reads from the compressed input
97 are done through this function. All reads are big endian */
99 static unsigned get_bits(bunzip_data *bd, int bits_wanted)
103 /* If we need to get more data from the byte buffer, do so. (Loop getting
104 one byte at a time to enforce endianness and avoid unaligned access.) */
106 while (bd->inbufBitCount < bits_wanted) {
108 /* If we need to read more data from file into byte buffer, do so */
110 if (bd->inbufPos == bd->inbufCount) {
111 /* if "no input fd" case: in_fd == -1, read fails, we jump */
112 bd->inbufCount = read(bd->in_fd, bd->inbuf, IOBUF_SIZE);
113 if (bd->inbufCount <= 0)
114 longjmp(bd->jmpbuf, RETVAL_UNEXPECTED_INPUT_EOF);
118 /* Avoid 32-bit overflow (dump bit buffer to top of output) */
120 if (bd->inbufBitCount >= 24) {
121 bits = bd->inbufBits & ((1 << bd->inbufBitCount) - 1);
122 bits_wanted -= bd->inbufBitCount;
123 bits <<= bits_wanted;
124 bd->inbufBitCount = 0;
127 /* Grab next 8 bits of input from buffer. */
129 bd->inbufBits = (bd->inbufBits << 8) | bd->inbuf[bd->inbufPos++];
130 bd->inbufBitCount += 8;
133 /* Calculate result */
135 bd->inbufBitCount -= bits_wanted;
136 bits |= (bd->inbufBits >> bd->inbufBitCount) & ((1 << bits_wanted) - 1);
141 /* Unpacks the next block and sets up for the inverse burrows-wheeler step. */
143 static int get_next_block(bunzip_data *bd)
145 struct group_data *hufGroup;
146 int dbufCount, nextSym, dbufSize, groupCount, *base, *limit, selector,
147 i, j, k, t, runPos, symCount, symTotal, nSelectors, byteCount[256];
148 unsigned char uc, symToByte[256], mtfSymbol[256], *selectors;
149 unsigned *dbuf, origPtr;
152 dbufSize = bd->dbufSize;
153 selectors = bd->selectors;
155 /* Reset longjmp I/O error handling */
157 i = setjmp(bd->jmpbuf);
160 /* Read in header signature and CRC, then validate signature.
161 (last block signature means CRC is for whole file, return now) */
163 i = get_bits(bd, 24);
164 j = get_bits(bd, 24);
165 bd->headerCRC = get_bits(bd, 32);
166 if ((i == 0x177245) && (j == 0x385090)) return RETVAL_LAST_BLOCK;
167 if ((i != 0x314159) || (j != 0x265359)) return RETVAL_NOT_BZIP_DATA;
169 /* We can add support for blockRandomised if anybody complains. There was
170 some code for this in busybox 1.0.0-pre3, but nobody ever noticed that
171 it didn't actually work. */
173 if (get_bits(bd, 1)) return RETVAL_OBSOLETE_INPUT;
174 origPtr = get_bits(bd, 24);
175 if (origPtr > dbufSize) return RETVAL_DATA_ERROR;
177 /* mapping table: if some byte values are never used (encoding things
178 like ascii text), the compression code removes the gaps to have fewer
179 symbols to deal with, and writes a sparse bitfield indicating which
180 values were present. We make a translation table to convert the symbols
181 back to the corresponding bytes. */
183 t = get_bits(bd, 16);
185 for (i = 0; i < 16; i++) {
186 if (t & (1 << (15-i))) {
187 k = get_bits(bd, 16);
188 for (j = 0; j < 16; j++)
189 if (k & (1 << (15-j)))
190 symToByte[symTotal++] = (16*i) + j;
194 /* How many different Huffman coding groups does this block use? */
196 groupCount = get_bits(bd, 3);
197 if (groupCount < 2 || groupCount > MAX_GROUPS)
198 return RETVAL_DATA_ERROR;
200 /* nSelectors: Every GROUP_SIZE many symbols we select a new Huffman coding
201 group. Read in the group selector list, which is stored as MTF encoded
202 bit runs. (MTF=Move To Front, as each value is used it's moved to the
203 start of the list.) */
205 nSelectors = get_bits(bd, 15);
206 if (!nSelectors) return RETVAL_DATA_ERROR;
207 for (i = 0; i < groupCount; i++) mtfSymbol[i] = i;
208 for (i = 0; i < nSelectors; i++) {
212 for (j = 0; get_bits(bd, 1); j++)
213 if (j >= groupCount) return RETVAL_DATA_ERROR;
215 /* Decode MTF to get the next selector */
218 for (;j;j--) mtfSymbol[j] = mtfSymbol[j-1];
219 mtfSymbol[0] = selectors[i] = uc;
222 /* Read the Huffman coding tables for each group, which code for symTotal
223 literal symbols, plus two run symbols (RUNA, RUNB) */
225 symCount = symTotal + 2;
226 for (j = 0; j < groupCount; j++) {
227 unsigned char length[MAX_SYMBOLS], temp[MAX_HUFCODE_BITS+1];
228 int minLen, maxLen, pp;
230 /* Read Huffman code lengths for each symbol. They're stored in
231 a way similar to mtf; record a starting value for the first symbol,
232 and an offset from the previous value for everys symbol after that.
233 (Subtracting 1 before the loop and then adding it back at the end is
234 an optimization that makes the test inside the loop simpler: symbol
235 length 0 becomes negative, so an unsigned inequality catches it.) */
237 t = get_bits(bd, 5) - 1;
238 for (i = 0; i < symCount; i++) {
240 if ((unsigned)t > (MAX_HUFCODE_BITS-1))
241 return RETVAL_DATA_ERROR;
243 /* If first bit is 0, stop. Else second bit indicates whether
244 to increment or decrement the value. Optimization: grab 2
245 bits and unget the second if the first was 0. */
253 /* Add one if second bit 1, else subtract 1. Avoids if/else */
255 t += (((k+1) & 2) - 1);
258 /* Correct for the initial -1, to get the final symbol length */
263 /* Find largest and smallest lengths in this group */
265 minLen = maxLen = length[0];
266 for (i = 1; i < symCount; i++) {
267 if (length[i] > maxLen) maxLen = length[i];
268 else if (length[i] < minLen) minLen = length[i];
271 /* Calculate permute[], base[], and limit[] tables from length[].
273 * permute[] is the lookup table for converting Huffman coded symbols
274 * into decoded symbols. base[] is the amount to subtract from the
275 * value of a Huffman symbol of a given length when using permute[].
277 * limit[] indicates the largest numerical value a symbol with a given
278 * number of bits can have. This is how the Huffman codes can vary in
279 * length: each code with a value>limit[length] needs another bit.
282 hufGroup = bd->groups + j;
283 hufGroup->minLen = minLen;
284 hufGroup->maxLen = maxLen;
286 /* Note that minLen can't be smaller than 1, so we adjust the base
287 and limit array pointers so we're not always wasting the first
288 entry. We do this again when using them (during symbol decoding).*/
290 base = hufGroup->base - 1;
291 limit = hufGroup->limit - 1;
293 /* Calculate permute[]. Concurently, initialize temp[] and limit[]. */
296 for (i = minLen; i <= maxLen; i++) {
297 temp[i] = limit[i] = 0;
298 for (t = 0; t < symCount; t++)
300 hufGroup->permute[pp++] = t;
303 /* Count symbols coded for at each bit length */
305 for (i = 0; i < symCount; i++) temp[length[i]]++;
307 /* Calculate limit[] (the largest symbol-coding value at each bit
308 * length, which is (previous limit<<1)+symbols at this level), and
309 * base[] (number of symbols to ignore at each bit length, which is
310 * limit minus the cumulative count of symbols coded for already). */
313 for (i = minLen; i < maxLen; i++) {
316 /* We read the largest possible symbol size and then unget bits
317 after determining how many we need, and those extra bits could
318 be set to anything. (They're noise from future symbols.) At
319 each level we're really only interested in the first few bits,
320 so here we set all the trailing to-be-ignored bits to 1 so they
321 don't affect the value>limit[length] comparison. */
323 limit[i] = (pp << (maxLen - i)) - 1;
328 limit[maxLen+1] = INT_MAX; /* Sentinel value for reading next sym. */
329 limit[maxLen] = pp + temp[maxLen] - 1;
333 /* We've finished reading and digesting the block header. Now read this
334 block's Huffman coded symbols from the file and undo the Huffman coding
335 and run length encoding, saving the result into dbuf[dbufCount++] = uc */
337 /* Initialize symbol occurrence counters and symbol Move To Front table */
339 memset(byteCount, 0, sizeof(byteCount)); /* smaller, maybe slower? */
340 for (i = 0; i < 256; i++) {
342 mtfSymbol[i] = (unsigned char)i;
345 /* Loop through compressed symbols. */
347 runPos = dbufCount = selector = 0;
350 /* fetch next Huffman coding group from list. */
352 symCount = GROUP_SIZE - 1;
353 if (selector >= nSelectors) return RETVAL_DATA_ERROR;
354 hufGroup = bd->groups + selectors[selector++];
355 base = hufGroup->base - 1;
356 limit = hufGroup->limit - 1;
359 /* Read next Huffman-coded symbol. */
361 /* Note: It is far cheaper to read maxLen bits and back up than it is
362 to read minLen bits and then an additional bit at a time, testing
363 as we go. Because there is a trailing last block (with file CRC),
364 there is no danger of the overread causing an unexpected EOF for a
365 valid compressed file. As a further optimization, we do the read
366 inline (falling back to a call to get_bits if the buffer runs
367 dry). The following (up to got_huff_bits:) is equivalent to
368 j = get_bits(bd, hufGroup->maxLen);
371 while (bd->inbufBitCount < hufGroup->maxLen) {
372 if (bd->inbufPos == bd->inbufCount) {
373 j = get_bits(bd, hufGroup->maxLen);
376 bd->inbufBits = (bd->inbufBits << 8) | bd->inbuf[bd->inbufPos++];
377 bd->inbufBitCount += 8;
379 bd->inbufBitCount -= hufGroup->maxLen;
380 j = (bd->inbufBits >> bd->inbufBitCount) & ((1 << hufGroup->maxLen) - 1);
384 /* Figure how how many bits are in next symbol and unget extras */
386 i = hufGroup->minLen;
387 while (j > limit[i]) ++i;
388 bd->inbufBitCount += (hufGroup->maxLen - i);
390 /* Huffman decode value to get nextSym (with bounds checking) */
392 if (i > hufGroup->maxLen)
393 return RETVAL_DATA_ERROR;
394 j = (j >> (hufGroup->maxLen - i)) - base[i];
395 if ((unsigned)j >= MAX_SYMBOLS)
396 return RETVAL_DATA_ERROR;
397 nextSym = hufGroup->permute[j];
399 /* We have now decoded the symbol, which indicates either a new literal
400 byte, or a repeated run of the most recent literal byte. First,
401 check if nextSym indicates a repeated run, and if so loop collecting
402 how many times to repeat the last literal. */
404 if ((unsigned)nextSym <= SYMBOL_RUNB) { /* RUNA or RUNB */
406 /* If this is the start of a new run, zero out counter */
413 /* Neat trick that saves 1 symbol: instead of or-ing 0 or 1 at
414 each bit position, add 1 or 2 instead. For example,
415 1011 is 1<<0 + 1<<1 + 2<<2. 1010 is 2<<0 + 2<<1 + 1<<2.
416 You can make any bit pattern that way using 1 less symbol than
417 the basic or 0/1 method (except all bits 0, which would use no
418 symbols, but a run of length 0 doesn't mean anything in this
419 context). Thus space is saved. */
421 t += (runPos << nextSym); /* +runPos if RUNA; +2*runPos if RUNB */
422 if (runPos < dbufSize) runPos <<= 1;
423 goto end_of_huffman_loop;
426 /* When we hit the first non-run symbol after a run, we now know
427 how many times to repeat the last literal, so append that many
428 copies to our buffer of decoded symbols (dbuf) now. (The last
429 literal used is the one at the head of the mtfSymbol array.) */
433 if (dbufCount + t >= dbufSize) return RETVAL_DATA_ERROR;
435 uc = symToByte[mtfSymbol[0]];
437 while (t--) dbuf[dbufCount++] = uc;
440 /* Is this the terminating symbol? */
442 if (nextSym > symTotal) break;
444 /* At this point, nextSym indicates a new literal character. Subtract
445 one to get the position in the MTF array at which this literal is
446 currently to be found. (Note that the result can't be -1 or 0,
447 because 0 and 1 are RUNA and RUNB. But another instance of the
448 first symbol in the mtf array, position 0, would have been handled
449 as part of a run above. Therefore 1 unused mtf position minus
450 2 non-literal nextSym values equals -1.) */
452 if (dbufCount >= dbufSize) return RETVAL_DATA_ERROR;
456 /* Adjust the MTF array. Since we typically expect to move only a
457 * small number of symbols, and are bound by 256 in any case, using
458 * memmove here would typically be bigger and slower due to function
459 * call overhead and other assorted setup costs. */
462 mtfSymbol[i] = mtfSymbol[i-1];
467 /* We have our literal byte. Save it into dbuf. */
470 dbuf[dbufCount++] = (unsigned)uc;
472 /* Skip group initialization if we're not done with this group. Done
473 * this way to avoid compiler warning. */
476 if (symCount--) goto continue_this_group;
479 /* At this point, we've read all the Huffman-coded symbols (and repeated
480 runs) for this block from the input stream, and decoded them into the
481 intermediate buffer. There are dbufCount many decoded bytes in dbuf[].
482 Now undo the Burrows-Wheeler transform on dbuf.
483 See http://dogma.net/markn/articles/bwt/bwt.htm
486 /* Turn byteCount into cumulative occurrence counts of 0 to n-1. */
489 for (i = 0; i < 256; i++) {
490 k = j + byteCount[i];
495 /* Figure out what order dbuf would be in if we sorted it. */
497 for (i = 0; i < dbufCount; i++) {
498 uc = (unsigned char)(dbuf[i] & 0xff);
499 dbuf[byteCount[uc]] |= (i << 8);
503 /* Decode first byte by hand to initialize "previous" byte. Note that it
504 doesn't get output, and if the first three characters are identical
505 it doesn't qualify as a run (hence writeRunCountdown=5). */
508 if (origPtr >= dbufCount) return RETVAL_DATA_ERROR;
509 bd->writePos = dbuf[origPtr];
510 bd->writeCurrent = (unsigned char)(bd->writePos & 0xff);
512 bd->writeRunCountdown = 5;
514 bd->writeCount = dbufCount;
519 /* Undo burrows-wheeler transform on intermediate buffer to produce output.
520 If start_bunzip was initialized with out_fd=-1, then up to len bytes of
521 data are written to outbuf. Return value is number of bytes written or
522 error (all errors are negative numbers). If out_fd!=-1, outbuf and len
523 are ignored, data is written to out_fd and return is RETVAL_OK or error.
526 int read_bunzip(bunzip_data *bd, char *outbuf, int len)
528 const unsigned *dbuf;
529 int pos, current, previous, gotcount;
531 /* If last read was short due to end of file, return last block now */
532 if (bd->writeCount < 0) return bd->writeCount;
537 current = bd->writeCurrent;
539 /* We will always have pending decoded data to write into the output
540 buffer unless this is the very first call (in which case we haven't
541 Huffman-decoded a block into the intermediate buffer yet). */
543 if (bd->writeCopies) {
545 /* Inside the loop, writeCopies means extra copies (beyond 1) */
549 /* Loop outputting bytes */
553 /* If the output buffer is full, snapshot state and return */
555 if (gotcount >= len) {
557 bd->writeCurrent = current;
562 /* Write next byte into output buffer, updating CRC */
564 outbuf[gotcount++] = current;
565 bd->writeCRC = (bd->writeCRC << 8)
566 ^ bd->crc32Table[(bd->writeCRC >> 24) ^ current];
568 /* Loop now if we're outputting multiple copies of this byte */
570 if (bd->writeCopies) {
575 if (!bd->writeCount--) break;
576 /* Follow sequence vector to undo Burrows-Wheeler transform */
579 current = pos & 0xff;
582 /* After 3 consecutive copies of the same byte, the 4th
583 * is a repeat count. We count down from 4 instead
584 * of counting up because testing for non-zero is faster */
586 if (--bd->writeRunCountdown) {
587 if (current != previous)
588 bd->writeRunCountdown = 4;
591 /* We have a repeated run, this byte indicates the count */
593 bd->writeCopies = current;
595 bd->writeRunCountdown = 5;
597 /* Sometimes there are just 3 bytes (run length 0) */
599 if (!bd->writeCopies) goto decode_next_byte;
601 /* Subtract the 1 copy we'd output anyway to get extras */
607 /* Decompression of this block completed successfully */
609 bd->writeCRC = ~bd->writeCRC;
610 bd->totalCRC = ((bd->totalCRC << 1) | (bd->totalCRC >> 31)) ^ bd->writeCRC;
612 /* If this block had a CRC error, force file level CRC error. */
614 if (bd->writeCRC != bd->headerCRC) {
615 bd->totalCRC = bd->headerCRC + 1;
616 return RETVAL_LAST_BLOCK;
620 /* Refill the intermediate buffer by Huffman-decoding next block of input */
621 /* (previous is just a convenient unused temp variable here) */
623 previous = get_next_block(bd);
625 bd->writeCount = previous;
626 return (previous != RETVAL_LAST_BLOCK) ? previous : gotcount;
630 current = bd->writeCurrent;
631 goto decode_next_byte;
635 /* Allocate the structure, read file header. If in_fd==-1, inbuf must contain
636 a complete bunzip file (len bytes long). If in_fd!=-1, inbuf and len are
637 ignored, and data is read from file handle into temporary buffer. */
639 /* Because bunzip2 is used for help text unpacking, and because bb_show_usage()
640 should work for NOFORK applets too, we must be extremely careful to not leak
643 int start_bunzip(bunzip_data **bdp, int in_fd, const unsigned char *inbuf,
649 BZh0 = ('B' << 24) + ('Z' << 16) + ('h' << 8) + '0'
652 /* Figure out how much data to allocate */
654 i = sizeof(bunzip_data);
655 if (in_fd != -1) i += IOBUF_SIZE;
657 /* Allocate bunzip_data. Most fields initialize to zero. */
659 bd = *bdp = xzalloc(i);
661 /* Setup input buffer */
665 /* in this case, bd->inbuf is read-only */
666 bd->inbuf = (void*)inbuf; /* cast away const-ness */
667 bd->inbufCount = len;
669 bd->inbuf = (unsigned char *)(bd + 1);
671 /* Init the CRC32 table (big endian) */
673 crc32_filltable(bd->crc32Table, 1);
675 /* Setup for I/O error handling via longjmp */
677 i = setjmp(bd->jmpbuf);
680 /* Ensure that file starts with "BZh['1'-'9']." */
682 i = get_bits(bd, 32);
683 if ((unsigned)(i - BZh0 - 1) >= 9) return RETVAL_NOT_BZIP_DATA;
685 /* Fourth byte (ascii '1'-'9'), indicates block size in units of 100k of
686 uncompressed data. Allocate intermediate buffer for block. */
688 bd->dbufSize = 100000 * (i - BZh0);
690 /* Cannot use xmalloc - may leak bd in NOFORK case! */
691 bd->dbuf = malloc_or_warn(bd->dbufSize * sizeof(int));
699 void dealloc_bunzip(bunzip_data *bd)
706 /* Decompress src_fd to dst_fd. Stops at end of bzip data, not end of file. */
708 USE_DESKTOP(long long) int
709 unpack_bz2_stream(int src_fd, int dst_fd)
711 USE_DESKTOP(long long total_written = 0;)
716 outbuf = xmalloc(IOBUF_SIZE);
717 i = start_bunzip(&bd, src_fd, NULL, 0);
720 i = read_bunzip(bd, outbuf, IOBUF_SIZE);
722 if (i != full_write(dst_fd, outbuf, i)) {
723 i = RETVAL_SHORT_WRITE;
726 USE_DESKTOP(total_written += i;)
730 /* Check CRC and release memory */
732 if (i == RETVAL_LAST_BLOCK) {
733 if (bd->headerCRC != bd->totalCRC) {
734 bb_error_msg("CRC error");
738 } else if (i == RETVAL_SHORT_WRITE) {
739 bb_error_msg("short write");
741 bb_error_msg("bunzip error %d", i);
746 return i ? i : USE_DESKTOP(total_written) + 0;
751 static char *const bunzip_errors[] = {
752 NULL, "Bad file checksum", "Not bzip data",
753 "Unexpected input EOF", "Unexpected output EOF", "Data error",
754 "Out of memory", "Obsolete (pre 0.9.5) bzip format not supported"
757 /* Dumb little test thing, decompress stdin to stdout */
758 int main(int argc, char **argv)
760 int i = unpack_bz2_stream(0, 1);
764 fprintf(stderr,"%s\n", bunzip_errors[-i]);
765 else if (read(0, &c, 1))
766 fprintf(stderr,"Trailing garbage ignored\n");