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.
40 #include "unarchive.h"
42 /* Constants for Huffman coding */
44 #define GROUP_SIZE 50 /* 64 would have been more efficient */
45 #define MAX_HUFCODE_BITS 20 /* Longest Huffman code allowed */
46 #define MAX_SYMBOLS 258 /* 256 literals + RUNA + RUNB */
50 /* Status return values */
52 #define RETVAL_LAST_BLOCK (-1)
53 #define RETVAL_NOT_BZIP_DATA (-2)
54 #define RETVAL_UNEXPECTED_INPUT_EOF (-3)
55 #define RETVAL_UNEXPECTED_OUTPUT_EOF (-4)
56 #define RETVAL_DATA_ERROR (-5)
57 #define RETVAL_OUT_OF_MEMORY (-6)
58 #define RETVAL_OBSOLETE_INPUT (-7)
60 /* Other housekeeping constants */
61 #define IOBUF_SIZE 4096
63 /* This is what we know about each Huffman coding group */
65 /* We have an extra slot at the end of limit[] for a sentinal value. */
66 int limit[MAX_HUFCODE_BITS+1],base[MAX_HUFCODE_BITS],permute[MAX_SYMBOLS];
70 /* Structure holding all the housekeeping data, including IO buffers and
71 memory that persists between calls to bunzip */
74 /* State for interrupting output loop */
76 int writeCopies,writePos,writeRunCountdown,writeCount,writeCurrent;
78 /* I/O tracking data (file handles, buffers, positions, etc.) */
80 int in_fd,out_fd,inbufCount,inbufPos /*,outbufPos*/;
81 unsigned char *inbuf /*,*outbuf*/;
82 unsigned int inbufBitCount, inbufBits;
84 /* The CRC values stored in the block header and calculated from the data */
86 uint32_t headerCRC, totalCRC, writeCRC;
88 /* Intermediate buffer and its size (in bytes) */
90 unsigned int *dbuf, dbufSize;
92 /* These things are a bit too big to go on the stack */
94 unsigned char selectors[32768]; /* nSelectors=15 bits */
95 struct group_data groups[MAX_GROUPS]; /* Huffman coding tables */
97 /* For I/O error handling */
102 /* Return the next nnn bits of input. All reads from the compressed input
103 are done through this function. All reads are big endian */
105 static unsigned int get_bits(bunzip_data *bd, char bits_wanted)
109 /* If we need to get more data from the byte buffer, do so. (Loop getting
110 one byte at a time to enforce endianness and avoid unaligned access.) */
112 while (bd->inbufBitCount<bits_wanted) {
114 /* If we need to read more data from file into byte buffer, do so */
116 if(bd->inbufPos==bd->inbufCount) {
117 if((bd->inbufCount = read(bd->in_fd, bd->inbuf, IOBUF_SIZE)) <= 0)
118 longjmp(bd->jmpbuf,RETVAL_UNEXPECTED_INPUT_EOF);
122 /* Avoid 32-bit overflow (dump bit buffer to top of output) */
124 if(bd->inbufBitCount>=24) {
125 bits=bd->inbufBits&((1<<bd->inbufBitCount)-1);
126 bits_wanted-=bd->inbufBitCount;
131 /* Grab next 8 bits of input from buffer. */
133 bd->inbufBits=(bd->inbufBits<<8)|bd->inbuf[bd->inbufPos++];
134 bd->inbufBitCount+=8;
137 /* Calculate result */
139 bd->inbufBitCount-=bits_wanted;
140 bits|=(bd->inbufBits>>bd->inbufBitCount)&((1<<bits_wanted)-1);
145 /* Unpacks the next block and sets up for the inverse burrows-wheeler step. */
147 static int get_next_block(bunzip_data *bd)
149 struct group_data *hufGroup;
150 int dbufCount,nextSym,dbufSize,groupCount,*base,*limit,selector,
151 i,j,k,t,runPos,symCount,symTotal,nSelectors,byteCount[256];
152 unsigned char uc, symToByte[256], mtfSymbol[256], *selectors;
153 unsigned int *dbuf,origPtr;
156 dbufSize=bd->dbufSize;
157 selectors=bd->selectors;
159 /* Reset longjmp I/O error handling */
161 i=setjmp(bd->jmpbuf);
164 /* Read in header signature and CRC, then validate signature.
165 (last block signature means CRC is for whole file, return now) */
169 bd->headerCRC=get_bits(bd,32);
170 if ((i == 0x177245) && (j == 0x385090)) return RETVAL_LAST_BLOCK;
171 if ((i != 0x314159) || (j != 0x265359)) return RETVAL_NOT_BZIP_DATA;
173 /* We can add support for blockRandomised if anybody complains. There was
174 some code for this in busybox 1.0.0-pre3, but nobody ever noticed that
175 it didn't actually work. */
177 if(get_bits(bd,1)) return RETVAL_OBSOLETE_INPUT;
178 if((origPtr=get_bits(bd,24)) > dbufSize) return RETVAL_DATA_ERROR;
180 /* mapping table: if some byte values are never used (encoding things
181 like ascii text), the compression code removes the gaps to have fewer
182 symbols to deal with, and writes a sparse bitfield indicating which
183 values were present. We make a translation table to convert the symbols
184 back to the corresponding bytes. */
192 if(k&(1<<(15-j))) symToByte[symTotal++]=(16*i)+j;
196 /* How many different Huffman coding groups does this block use? */
198 groupCount=get_bits(bd,3);
199 if (groupCount<2 || groupCount>MAX_GROUPS) return RETVAL_DATA_ERROR;
201 /* nSelectors: Every GROUP_SIZE many symbols we select a new Huffman coding
202 group. Read in the group selector list, which is stored as MTF encoded
203 bit runs. (MTF=Move To Front, as each value is used it's moved to the
204 start of the list.) */
206 if(!(nSelectors=get_bits(bd, 15))) 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++) if (j>=groupCount) return RETVAL_DATA_ERROR;
214 /* Decode MTF to get the next selector */
217 for(;j;j--) mtfSymbol[j] = mtfSymbol[j-1];
218 mtfSymbol[0]=selectors[i]=uc;
221 /* Read the Huffman coding tables for each group, which code for symTotal
222 literal symbols, plus two run symbols (RUNA, RUNB) */
225 for (j=0; j<groupCount; j++) {
226 unsigned char length[MAX_SYMBOLS],temp[MAX_HUFCODE_BITS+1];
227 int minLen, maxLen, pp;
229 /* Read Huffman code lengths for each symbol. They're stored in
230 a way similar to mtf; record a starting value for the first symbol,
231 and an offset from the previous value for everys symbol after that.
232 (Subtracting 1 before the loop and then adding it back at the end is
233 an optimization that makes the test inside the loop simpler: symbol
234 length 0 becomes negative, so an unsigned inequality catches it.) */
237 for (i = 0; i < symCount; i++) {
239 if (((unsigned)t) > (MAX_HUFCODE_BITS-1))
240 return RETVAL_DATA_ERROR;
242 /* If first bit is 0, stop. Else second bit indicates whether
243 to increment or decrement the value. Optimization: grab 2
244 bits and unget the second if the first was 0. */
252 /* Add one if second bit 1, else subtract 1. Avoids if/else */
257 /* Correct for the initial -1, to get the final symbol length */
262 /* Find largest and smallest lengths in this group */
264 minLen=maxLen=length[0];
265 for(i = 1; i < symCount; i++) {
266 if(length[i] > maxLen) maxLen = length[i];
267 else if(length[i] < minLen) minLen = length[i];
270 /* Calculate permute[], base[], and limit[] tables from length[].
272 * permute[] is the lookup table for converting Huffman coded symbols
273 * into decoded symbols. base[] is the amount to subtract from the
274 * value of a Huffman symbol of a given length when using permute[].
276 * limit[] indicates the largest numerical value a symbol with a given
277 * number of bits can have. This is how the Huffman codes can vary in
278 * length: each code with a value>limit[length] needs another bit.
281 hufGroup=bd->groups+j;
282 hufGroup->minLen = minLen;
283 hufGroup->maxLen = maxLen;
285 /* Note that minLen can't be smaller than 1, so we adjust the base
286 and limit array pointers so we're not always wasting the first
287 entry. We do this again when using them (during symbol decoding).*/
289 base=hufGroup->base-1;
290 limit=hufGroup->limit-1;
292 /* Calculate permute[]. Concurently, initialize temp[] and limit[]. */
295 for(i=minLen;i<=maxLen;i++) {
297 for(t=0;t<symCount;t++)
298 if(length[t]==i) hufGroup->permute[pp++] = t;
301 /* Count symbols coded for at each bit length */
303 for (i=0;i<symCount;i++) temp[length[i]]++;
305 /* Calculate limit[] (the largest symbol-coding value at each bit
306 * length, which is (previous limit<<1)+symbols at this level), and
307 * base[] (number of symbols to ignore at each bit length, which is
308 * limit minus the cumulative count of symbols coded for already). */
311 for (i=minLen; i<maxLen; i++) {
314 /* We read the largest possible symbol size and then unget bits
315 after determining how many we need, and those extra bits could
316 be set to anything. (They're noise from future symbols.) At
317 each level we're really only interested in the first few bits,
318 so here we set all the trailing to-be-ignored bits to 1 so they
319 don't affect the value>limit[length] comparison. */
321 limit[i]= (pp << (maxLen - i)) - 1;
323 base[i+1]=pp-(t+=temp[i]);
325 limit[maxLen+1] = INT_MAX; /* Sentinal value for reading next sym. */
326 limit[maxLen]=pp+temp[maxLen]-1;
330 /* We've finished reading and digesting the block header. Now read this
331 block's Huffman coded symbols from the file and undo the Huffman coding
332 and run length encoding, saving the result into dbuf[dbufCount++]=uc */
334 /* Initialize symbol occurrence counters and symbol Move To Front table */
338 mtfSymbol[i]=(unsigned char)i;
341 /* Loop through compressed symbols. */
343 runPos=dbufCount=selector=0;
346 /* fetch next Huffman coding group from list. */
348 symCount=GROUP_SIZE-1;
349 if(selector>=nSelectors) return RETVAL_DATA_ERROR;
350 hufGroup=bd->groups+selectors[selector++];
351 base=hufGroup->base-1;
352 limit=hufGroup->limit-1;
355 /* Read next Huffman-coded symbol. */
357 /* Note: It is far cheaper to read maxLen bits and back up than it is
358 to read minLen bits and then an additional bit at a time, testing
359 as we go. Because there is a trailing last block (with file CRC),
360 there is no danger of the overread causing an unexpected EOF for a
361 valid compressed file. As a further optimization, we do the read
362 inline (falling back to a call to get_bits if the buffer runs
363 dry). The following (up to got_huff_bits:) is equivalent to
364 j=get_bits(bd,hufGroup->maxLen);
367 while (bd->inbufBitCount<hufGroup->maxLen) {
368 if(bd->inbufPos==bd->inbufCount) {
369 j = get_bits(bd,hufGroup->maxLen);
372 bd->inbufBits=(bd->inbufBits<<8)|bd->inbuf[bd->inbufPos++];
373 bd->inbufBitCount+=8;
375 bd->inbufBitCount-=hufGroup->maxLen;
376 j = (bd->inbufBits>>bd->inbufBitCount)&((1<<hufGroup->maxLen)-1);
380 /* Figure how how many bits are in next symbol and unget extras */
383 while(j>limit[i]) ++i;
384 bd->inbufBitCount += (hufGroup->maxLen - i);
386 /* Huffman decode value to get nextSym (with bounds checking) */
388 if ((i > hufGroup->maxLen)
389 || (((unsigned)(j=(j>>(hufGroup->maxLen-i))-base[i]))
391 return RETVAL_DATA_ERROR;
392 nextSym = hufGroup->permute[j];
394 /* We have now decoded the symbol, which indicates either a new literal
395 byte, or a repeated run of the most recent literal byte. First,
396 check if nextSym indicates a repeated run, and if so loop collecting
397 how many times to repeat the last literal. */
399 if (((unsigned)nextSym) <= SYMBOL_RUNB) { /* RUNA or RUNB */
401 /* If this is the start of a new run, zero out counter */
408 /* Neat trick that saves 1 symbol: instead of or-ing 0 or 1 at
409 each bit position, add 1 or 2 instead. For example,
410 1011 is 1<<0 + 1<<1 + 2<<2. 1010 is 2<<0 + 2<<1 + 1<<2.
411 You can make any bit pattern that way using 1 less symbol than
412 the basic or 0/1 method (except all bits 0, which would use no
413 symbols, but a run of length 0 doesn't mean anything in this
414 context). Thus space is saved. */
416 t += (runPos << nextSym); /* +runPos if RUNA; +2*runPos if RUNB */
417 if(runPos < dbufSize) runPos <<= 1;
418 goto end_of_huffman_loop;
421 /* When we hit the first non-run symbol after a run, we now know
422 how many times to repeat the last literal, so append that many
423 copies to our buffer of decoded symbols (dbuf) now. (The last
424 literal used is the one at the head of the mtfSymbol array.) */
428 if(dbufCount+t>=dbufSize) return RETVAL_DATA_ERROR;
430 uc = symToByte[mtfSymbol[0]];
432 while(t--) dbuf[dbufCount++]=uc;
435 /* Is this the terminating symbol? */
437 if(nextSym>symTotal) break;
439 /* At this point, nextSym indicates a new literal character. Subtract
440 one to get the position in the MTF array at which this literal is
441 currently to be found. (Note that the result can't be -1 or 0,
442 because 0 and 1 are RUNA and RUNB. But another instance of the
443 first symbol in the mtf array, position 0, would have been handled
444 as part of a run above. Therefore 1 unused mtf position minus
445 2 non-literal nextSym values equals -1.) */
447 if(dbufCount>=dbufSize) return RETVAL_DATA_ERROR;
451 /* Adjust the MTF array. Since we typically expect to move only a
452 * small number of symbols, and are bound by 256 in any case, using
453 * memmove here would typically be bigger and slower due to function
454 * call overhead and other assorted setup costs. */
457 mtfSymbol[i] = mtfSymbol[i-1];
462 /* We have our literal byte. Save it into dbuf. */
465 dbuf[dbufCount++] = (unsigned int)uc;
467 /* Skip group initialization if we're not done with this group. Done
468 * this way to avoid compiler warning. */
471 if(symCount--) goto continue_this_group;
474 /* At this point, we've read all the Huffman-coded symbols (and repeated
475 runs) for this block from the input stream, and decoded them into the
476 intermediate buffer. There are dbufCount many decoded bytes in dbuf[].
477 Now undo the Burrows-Wheeler transform on dbuf.
478 See http://dogma.net/markn/articles/bwt/bwt.htm
481 /* Turn byteCount into cumulative occurrence counts of 0 to n-1. */
490 /* Figure out what order dbuf would be in if we sorted it. */
492 for (i=0;i<dbufCount;i++) {
493 uc=(unsigned char)(dbuf[i] & 0xff);
494 dbuf[byteCount[uc]] |= (i << 8);
498 /* Decode first byte by hand to initialize "previous" byte. Note that it
499 doesn't get output, and if the first three characters are identical
500 it doesn't qualify as a run (hence writeRunCountdown=5). */
503 if(origPtr>=dbufCount) return RETVAL_DATA_ERROR;
504 bd->writePos=dbuf[origPtr];
505 bd->writeCurrent=(unsigned char)(bd->writePos&0xff);
507 bd->writeRunCountdown=5;
509 bd->writeCount=dbufCount;
514 /* Undo burrows-wheeler transform on intermediate buffer to produce output.
515 If start_bunzip was initialized with out_fd=-1, then up to len bytes of
516 data are written to outbuf. Return value is number of bytes written or
517 error (all errors are negative numbers). If out_fd!=-1, outbuf and len
518 are ignored, data is written to out_fd and return is RETVAL_OK or error.
521 static int read_bunzip(bunzip_data *bd, char *outbuf, int len)
523 const unsigned int *dbuf;
524 int pos,current,previous,gotcount;
526 /* If last read was short due to end of file, return last block now */
527 if(bd->writeCount<0) return bd->writeCount;
532 current=bd->writeCurrent;
534 /* We will always have pending decoded data to write into the output
535 buffer unless this is the very first call (in which case we haven't
536 Huffman-decoded a block into the intermediate buffer yet). */
538 if (bd->writeCopies) {
540 /* Inside the loop, writeCopies means extra copies (beyond 1) */
544 /* Loop outputting bytes */
548 /* If the output buffer is full, snapshot state and return */
550 if(gotcount >= len) {
552 bd->writeCurrent=current;
557 /* Write next byte into output buffer, updating CRC */
559 outbuf[gotcount++] = current;
560 bd->writeCRC=(((bd->writeCRC)<<8)
561 ^bd->crc32Table[((bd->writeCRC)>>24)^current]);
563 /* Loop now if we're outputting multiple copies of this byte */
565 if (bd->writeCopies) {
570 if (!bd->writeCount--) break;
571 /* Follow sequence vector to undo Burrows-Wheeler transform */
577 /* After 3 consecutive copies of the same byte, the 4th is a repeat
578 count. We count down from 4 instead
579 * of counting up because testing for non-zero is faster */
581 if(--bd->writeRunCountdown) {
582 if(current!=previous) bd->writeRunCountdown=4;
585 /* We have a repeated run, this byte indicates the count */
587 bd->writeCopies=current;
589 bd->writeRunCountdown=5;
591 /* Sometimes there are just 3 bytes (run length 0) */
593 if(!bd->writeCopies) goto decode_next_byte;
595 /* Subtract the 1 copy we'd output anyway to get extras */
601 /* Decompression of this block completed successfully */
603 bd->writeCRC=~bd->writeCRC;
604 bd->totalCRC=((bd->totalCRC<<1) | (bd->totalCRC>>31)) ^ bd->writeCRC;
606 /* If this block had a CRC error, force file level CRC error. */
608 if(bd->writeCRC!=bd->headerCRC) {
609 bd->totalCRC=bd->headerCRC+1;
610 return RETVAL_LAST_BLOCK;
614 /* Refill the intermediate buffer by Huffman-decoding next block of input */
615 /* (previous is just a convenient unused temp variable here) */
617 previous=get_next_block(bd);
619 bd->writeCount=previous;
620 return (previous!=RETVAL_LAST_BLOCK) ? previous : gotcount;
624 current=bd->writeCurrent;
625 goto decode_next_byte;
628 /* Allocate the structure, read file header. If in_fd==-1, inbuf must contain
629 a complete bunzip file (len bytes long). If in_fd!=-1, inbuf and len are
630 ignored, and data is read from file handle into temporary buffer. */
632 static int start_bunzip(bunzip_data **bdp, int in_fd, unsigned char *inbuf,
637 const unsigned int BZh0=(((unsigned int)'B')<<24)+(((unsigned int)'Z')<<16)
638 +(((unsigned int)'h')<<8)+(unsigned int)'0';
640 /* Figure out how much data to allocate */
642 i=sizeof(bunzip_data);
643 if(in_fd!=-1) i+=IOBUF_SIZE;
645 /* Allocate bunzip_data. Most fields initialize to zero. */
648 memset(bd,0,sizeof(bunzip_data));
650 /* Setup input buffer */
652 if(-1==(bd->in_fd=in_fd)) {
655 } else bd->inbuf=(unsigned char *)(bd+1);
657 /* Init the CRC32 table (big endian) */
659 bd->crc32Table = bb_crc32_filltable(1);
661 /* Setup for I/O error handling via longjmp */
663 i=setjmp(bd->jmpbuf);
666 /* Ensure that file starts with "BZh['1'-'9']." */
669 if (((unsigned int)(i-BZh0-1)) >= 9) return RETVAL_NOT_BZIP_DATA;
671 /* Fourth byte (ascii '1'-'9'), indicates block size in units of 100k of
672 uncompressed data. Allocate intermediate buffer for block. */
674 bd->dbufSize=100000*(i-BZh0);
676 bd->dbuf=xmalloc(bd->dbufSize * sizeof(int));
680 /* Example usage: decompress src_fd to dst_fd. (Stops at end of bzip data,
683 int uncompressStream(int src_fd, int dst_fd)
689 outbuf=xmalloc(IOBUF_SIZE);
690 if(!(i=start_bunzip(&bd,src_fd,0,0))) {
692 if((i=read_bunzip(bd,outbuf,IOBUF_SIZE)) <= 0) break;
693 if(i!=write(dst_fd,outbuf,i)) {
694 i=RETVAL_UNEXPECTED_OUTPUT_EOF;
700 /* Check CRC and release memory */
702 if(i==RETVAL_LAST_BLOCK) {
703 if (bd->headerCRC!=bd->totalCRC) {
704 bb_error_msg("Data integrity error when decompressing.");
708 } else if (i==RETVAL_UNEXPECTED_OUTPUT_EOF) {
709 bb_error_msg("Compressed file ends unexpectedly");
711 bb_error_msg("Decompression failed");
722 static char * const bunzip_errors[]={NULL,"Bad file checksum","Not bzip data",
723 "Unexpected input EOF","Unexpected output EOF","Data error",
724 "Out of memory","Obsolete (pre 0.9.5) bzip format not supported."};
726 /* Dumb little test thing, decompress stdin to stdout */
727 int main(int argc, char *argv[])
729 int i=uncompressStream(0,1);
732 if(i) fprintf(stderr,"%s\n", bunzip_errors[-i]);
733 else if(read(0,&c,1)) fprintf(stderr,"Trailing garbage ignored\n");