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_UNEXPECTED_OUTPUT_EOF (-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 sentinal 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 */
66 /* State for interrupting output loop */
68 int writeCopies,writePos,writeRunCountdown,writeCount,writeCurrent;
70 /* I/O tracking data (file handles, buffers, positions, etc.) */
72 int in_fd,out_fd,inbufCount,inbufPos /*,outbufPos*/;
73 unsigned char *inbuf /*,*outbuf*/;
74 unsigned int inbufBitCount, inbufBits;
76 /* The CRC values stored in the block header and calculated from the data */
78 uint32_t headerCRC, totalCRC, writeCRC;
80 /* Intermediate buffer and its size (in bytes) */
82 unsigned int *dbuf, dbufSize;
84 /* These things are a bit too big to go on the stack */
86 unsigned char selectors[32768]; /* nSelectors=15 bits */
87 struct group_data groups[MAX_GROUPS]; /* Huffman coding tables */
89 /* For I/O error handling */
94 /* Return the next nnn bits of input. All reads from the compressed input
95 are done through this function. All reads are big endian */
97 static unsigned int get_bits(bunzip_data *bd, char bits_wanted)
101 /* If we need to get more data from the byte buffer, do so. (Loop getting
102 one byte at a time to enforce endianness and avoid unaligned access.) */
104 while (bd->inbufBitCount<bits_wanted) {
106 /* If we need to read more data from file into byte buffer, do so */
108 if(bd->inbufPos==bd->inbufCount) {
109 if((bd->inbufCount = read(bd->in_fd, bd->inbuf, IOBUF_SIZE)) <= 0)
110 longjmp(bd->jmpbuf,RETVAL_UNEXPECTED_INPUT_EOF);
114 /* Avoid 32-bit overflow (dump bit buffer to top of output) */
116 if(bd->inbufBitCount>=24) {
117 bits=bd->inbufBits&((1<<bd->inbufBitCount)-1);
118 bits_wanted-=bd->inbufBitCount;
123 /* Grab next 8 bits of input from buffer. */
125 bd->inbufBits=(bd->inbufBits<<8)|bd->inbuf[bd->inbufPos++];
126 bd->inbufBitCount+=8;
129 /* Calculate result */
131 bd->inbufBitCount-=bits_wanted;
132 bits|=(bd->inbufBits>>bd->inbufBitCount)&((1<<bits_wanted)-1);
137 /* Unpacks the next block and sets up for the inverse burrows-wheeler step. */
139 static int get_next_block(bunzip_data *bd)
141 struct group_data *hufGroup;
142 int dbufCount,nextSym,dbufSize,groupCount,*base,*limit,selector,
143 i,j,k,t,runPos,symCount,symTotal,nSelectors,byteCount[256];
144 unsigned char uc, symToByte[256], mtfSymbol[256], *selectors;
145 unsigned int *dbuf,origPtr;
148 dbufSize=bd->dbufSize;
149 selectors=bd->selectors;
151 /* Reset longjmp I/O error handling */
153 i=setjmp(bd->jmpbuf);
156 /* Read in header signature and CRC, then validate signature.
157 (last block signature means CRC is for whole file, return now) */
161 bd->headerCRC=get_bits(bd,32);
162 if ((i == 0x177245) && (j == 0x385090)) return RETVAL_LAST_BLOCK;
163 if ((i != 0x314159) || (j != 0x265359)) return RETVAL_NOT_BZIP_DATA;
165 /* We can add support for blockRandomised if anybody complains. There was
166 some code for this in busybox 1.0.0-pre3, but nobody ever noticed that
167 it didn't actually work. */
169 if (get_bits(bd,1)) return RETVAL_OBSOLETE_INPUT;
170 if ((origPtr=get_bits(bd,24)) > dbufSize) return RETVAL_DATA_ERROR;
172 /* mapping table: if some byte values are never used (encoding things
173 like ascii text), the compression code removes the gaps to have fewer
174 symbols to deal with, and writes a sparse bitfield indicating which
175 values were present. We make a translation table to convert the symbols
176 back to the corresponding bytes. */
184 if(k&(1<<(15-j))) symToByte[symTotal++]=(16*i)+j;
188 /* How many different Huffman coding groups does this block use? */
190 groupCount=get_bits(bd,3);
191 if (groupCount<2 || groupCount>MAX_GROUPS) return RETVAL_DATA_ERROR;
193 /* nSelectors: Every GROUP_SIZE many symbols we select a new Huffman coding
194 group. Read in the group selector list, which is stored as MTF encoded
195 bit runs. (MTF=Move To Front, as each value is used it's moved to the
196 start of the list.) */
198 if(!(nSelectors=get_bits(bd, 15))) return RETVAL_DATA_ERROR;
199 for (i=0; i<groupCount; i++) mtfSymbol[i] = i;
200 for (i=0; i<nSelectors; i++) {
204 for (j=0;get_bits(bd,1);j++) if (j>=groupCount) return RETVAL_DATA_ERROR;
206 /* Decode MTF to get the next selector */
209 for (;j;j--) mtfSymbol[j] = mtfSymbol[j-1];
210 mtfSymbol[0]=selectors[i]=uc;
213 /* Read the Huffman coding tables for each group, which code for symTotal
214 literal symbols, plus two run symbols (RUNA, RUNB) */
217 for (j=0; j<groupCount; j++) {
218 unsigned char length[MAX_SYMBOLS],temp[MAX_HUFCODE_BITS+1];
219 int minLen, maxLen, pp;
221 /* Read Huffman code lengths for each symbol. They're stored in
222 a way similar to mtf; record a starting value for the first symbol,
223 and an offset from the previous value for everys symbol after that.
224 (Subtracting 1 before the loop and then adding it back at the end is
225 an optimization that makes the test inside the loop simpler: symbol
226 length 0 becomes negative, so an unsigned inequality catches it.) */
229 for (i = 0; i < symCount; i++) {
231 if (((unsigned)t) > (MAX_HUFCODE_BITS-1))
232 return RETVAL_DATA_ERROR;
234 /* If first bit is 0, stop. Else second bit indicates whether
235 to increment or decrement the value. Optimization: grab 2
236 bits and unget the second if the first was 0. */
244 /* Add one if second bit 1, else subtract 1. Avoids if/else */
249 /* Correct for the initial -1, to get the final symbol length */
254 /* Find largest and smallest lengths in this group */
256 minLen=maxLen=length[0];
257 for (i = 1; i < symCount; i++) {
258 if(length[i] > maxLen) maxLen = length[i];
259 else if(length[i] < minLen) minLen = length[i];
262 /* Calculate permute[], base[], and limit[] tables from length[].
264 * permute[] is the lookup table for converting Huffman coded symbols
265 * into decoded symbols. base[] is the amount to subtract from the
266 * value of a Huffman symbol of a given length when using permute[].
268 * limit[] indicates the largest numerical value a symbol with a given
269 * number of bits can have. This is how the Huffman codes can vary in
270 * length: each code with a value>limit[length] needs another bit.
273 hufGroup=bd->groups+j;
274 hufGroup->minLen = minLen;
275 hufGroup->maxLen = maxLen;
277 /* Note that minLen can't be smaller than 1, so we adjust the base
278 and limit array pointers so we're not always wasting the first
279 entry. We do this again when using them (during symbol decoding).*/
281 base=hufGroup->base-1;
282 limit=hufGroup->limit-1;
284 /* Calculate permute[]. Concurently, initialize temp[] and limit[]. */
287 for (i=minLen;i<=maxLen;i++) {
289 for (t=0;t<symCount;t++)
290 if(length[t]==i) hufGroup->permute[pp++] = t;
293 /* Count symbols coded for at each bit length */
295 for (i=0;i<symCount;i++) temp[length[i]]++;
297 /* Calculate limit[] (the largest symbol-coding value at each bit
298 * length, which is (previous limit<<1)+symbols at this level), and
299 * base[] (number of symbols to ignore at each bit length, which is
300 * limit minus the cumulative count of symbols coded for already). */
303 for (i=minLen; i<maxLen; i++) {
306 /* We read the largest possible symbol size and then unget bits
307 after determining how many we need, and those extra bits could
308 be set to anything. (They're noise from future symbols.) At
309 each level we're really only interested in the first few bits,
310 so here we set all the trailing to-be-ignored bits to 1 so they
311 don't affect the value>limit[length] comparison. */
313 limit[i]= (pp << (maxLen - i)) - 1;
315 base[i+1]=pp-(t+=temp[i]);
317 limit[maxLen+1] = INT_MAX; /* Sentinal value for reading next sym. */
318 limit[maxLen]=pp+temp[maxLen]-1;
322 /* We've finished reading and digesting the block header. Now read this
323 block's Huffman coded symbols from the file and undo the Huffman coding
324 and run length encoding, saving the result into dbuf[dbufCount++]=uc */
326 /* Initialize symbol occurrence counters and symbol Move To Front table */
328 for (i=0;i<256;i++) {
330 mtfSymbol[i]=(unsigned char)i;
333 /* Loop through compressed symbols. */
335 runPos=dbufCount=selector=0;
338 /* fetch next Huffman coding group from list. */
340 symCount=GROUP_SIZE-1;
341 if(selector>=nSelectors) return RETVAL_DATA_ERROR;
342 hufGroup=bd->groups+selectors[selector++];
343 base=hufGroup->base-1;
344 limit=hufGroup->limit-1;
347 /* Read next Huffman-coded symbol. */
349 /* Note: It is far cheaper to read maxLen bits and back up than it is
350 to read minLen bits and then an additional bit at a time, testing
351 as we go. Because there is a trailing last block (with file CRC),
352 there is no danger of the overread causing an unexpected EOF for a
353 valid compressed file. As a further optimization, we do the read
354 inline (falling back to a call to get_bits if the buffer runs
355 dry). The following (up to got_huff_bits:) is equivalent to
356 j=get_bits(bd,hufGroup->maxLen);
359 while (bd->inbufBitCount<hufGroup->maxLen) {
360 if(bd->inbufPos==bd->inbufCount) {
361 j = get_bits(bd,hufGroup->maxLen);
364 bd->inbufBits=(bd->inbufBits<<8)|bd->inbuf[bd->inbufPos++];
365 bd->inbufBitCount+=8;
367 bd->inbufBitCount-=hufGroup->maxLen;
368 j = (bd->inbufBits>>bd->inbufBitCount)&((1<<hufGroup->maxLen)-1);
372 /* Figure how how many bits are in next symbol and unget extras */
375 while (j>limit[i]) ++i;
376 bd->inbufBitCount += (hufGroup->maxLen - i);
378 /* Huffman decode value to get nextSym (with bounds checking) */
380 if ((i > hufGroup->maxLen)
381 || (((unsigned)(j=(j>>(hufGroup->maxLen-i))-base[i]))
383 return RETVAL_DATA_ERROR;
384 nextSym = hufGroup->permute[j];
386 /* We have now decoded the symbol, which indicates either a new literal
387 byte, or a repeated run of the most recent literal byte. First,
388 check if nextSym indicates a repeated run, and if so loop collecting
389 how many times to repeat the last literal. */
391 if (((unsigned)nextSym) <= SYMBOL_RUNB) { /* RUNA or RUNB */
393 /* If this is the start of a new run, zero out counter */
400 /* Neat trick that saves 1 symbol: instead of or-ing 0 or 1 at
401 each bit position, add 1 or 2 instead. For example,
402 1011 is 1<<0 + 1<<1 + 2<<2. 1010 is 2<<0 + 2<<1 + 1<<2.
403 You can make any bit pattern that way using 1 less symbol than
404 the basic or 0/1 method (except all bits 0, which would use no
405 symbols, but a run of length 0 doesn't mean anything in this
406 context). Thus space is saved. */
408 t += (runPos << nextSym); /* +runPos if RUNA; +2*runPos if RUNB */
409 if(runPos < dbufSize) runPos <<= 1;
410 goto end_of_huffman_loop;
413 /* When we hit the first non-run symbol after a run, we now know
414 how many times to repeat the last literal, so append that many
415 copies to our buffer of decoded symbols (dbuf) now. (The last
416 literal used is the one at the head of the mtfSymbol array.) */
420 if(dbufCount+t>=dbufSize) return RETVAL_DATA_ERROR;
422 uc = symToByte[mtfSymbol[0]];
424 while (t--) dbuf[dbufCount++]=uc;
427 /* Is this the terminating symbol? */
429 if(nextSym>symTotal) break;
431 /* At this point, nextSym indicates a new literal character. Subtract
432 one to get the position in the MTF array at which this literal is
433 currently to be found. (Note that the result can't be -1 or 0,
434 because 0 and 1 are RUNA and RUNB. But another instance of the
435 first symbol in the mtf array, position 0, would have been handled
436 as part of a run above. Therefore 1 unused mtf position minus
437 2 non-literal nextSym values equals -1.) */
439 if(dbufCount>=dbufSize) return RETVAL_DATA_ERROR;
443 /* Adjust the MTF array. Since we typically expect to move only a
444 * small number of symbols, and are bound by 256 in any case, using
445 * memmove here would typically be bigger and slower due to function
446 * call overhead and other assorted setup costs. */
449 mtfSymbol[i] = mtfSymbol[i-1];
454 /* We have our literal byte. Save it into dbuf. */
457 dbuf[dbufCount++] = (unsigned int)uc;
459 /* Skip group initialization if we're not done with this group. Done
460 * this way to avoid compiler warning. */
463 if(symCount--) goto continue_this_group;
466 /* At this point, we've read all the Huffman-coded symbols (and repeated
467 runs) for this block from the input stream, and decoded them into the
468 intermediate buffer. There are dbufCount many decoded bytes in dbuf[].
469 Now undo the Burrows-Wheeler transform on dbuf.
470 See http://dogma.net/markn/articles/bwt/bwt.htm
473 /* Turn byteCount into cumulative occurrence counts of 0 to n-1. */
476 for (i=0;i<256;i++) {
482 /* Figure out what order dbuf would be in if we sorted it. */
484 for (i=0;i<dbufCount;i++) {
485 uc=(unsigned char)(dbuf[i] & 0xff);
486 dbuf[byteCount[uc]] |= (i << 8);
490 /* Decode first byte by hand to initialize "previous" byte. Note that it
491 doesn't get output, and if the first three characters are identical
492 it doesn't qualify as a run (hence writeRunCountdown=5). */
495 if(origPtr>=dbufCount) return RETVAL_DATA_ERROR;
496 bd->writePos=dbuf[origPtr];
497 bd->writeCurrent=(unsigned char)(bd->writePos&0xff);
499 bd->writeRunCountdown=5;
501 bd->writeCount=dbufCount;
506 /* Undo burrows-wheeler transform on intermediate buffer to produce output.
507 If start_bunzip was initialized with out_fd=-1, then up to len bytes of
508 data are written to outbuf. Return value is number of bytes written or
509 error (all errors are negative numbers). If out_fd!=-1, outbuf and len
510 are ignored, data is written to out_fd and return is RETVAL_OK or error.
513 static int read_bunzip(bunzip_data *bd, char *outbuf, int len)
515 const unsigned int *dbuf;
516 int pos,current,previous,gotcount;
518 /* If last read was short due to end of file, return last block now */
519 if(bd->writeCount<0) return bd->writeCount;
524 current=bd->writeCurrent;
526 /* We will always have pending decoded data to write into the output
527 buffer unless this is the very first call (in which case we haven't
528 Huffman-decoded a block into the intermediate buffer yet). */
530 if (bd->writeCopies) {
532 /* Inside the loop, writeCopies means extra copies (beyond 1) */
536 /* Loop outputting bytes */
540 /* If the output buffer is full, snapshot state and return */
542 if(gotcount >= len) {
544 bd->writeCurrent=current;
549 /* Write next byte into output buffer, updating CRC */
551 outbuf[gotcount++] = current;
552 bd->writeCRC=(((bd->writeCRC)<<8)
553 ^bd->crc32Table[((bd->writeCRC)>>24)^current]);
555 /* Loop now if we're outputting multiple copies of this byte */
557 if (bd->writeCopies) {
562 if (!bd->writeCount--) break;
563 /* Follow sequence vector to undo Burrows-Wheeler transform */
569 /* After 3 consecutive copies of the same byte, the 4th is a repeat
570 count. We count down from 4 instead
571 * of counting up because testing for non-zero is faster */
573 if(--bd->writeRunCountdown) {
574 if(current!=previous) bd->writeRunCountdown=4;
577 /* We have a repeated run, this byte indicates the count */
579 bd->writeCopies=current;
581 bd->writeRunCountdown=5;
583 /* Sometimes there are just 3 bytes (run length 0) */
585 if(!bd->writeCopies) goto decode_next_byte;
587 /* Subtract the 1 copy we'd output anyway to get extras */
593 /* Decompression of this block completed successfully */
595 bd->writeCRC=~bd->writeCRC;
596 bd->totalCRC=((bd->totalCRC<<1) | (bd->totalCRC>>31)) ^ bd->writeCRC;
598 /* If this block had a CRC error, force file level CRC error. */
600 if(bd->writeCRC!=bd->headerCRC) {
601 bd->totalCRC=bd->headerCRC+1;
602 return RETVAL_LAST_BLOCK;
606 /* Refill the intermediate buffer by Huffman-decoding next block of input */
607 /* (previous is just a convenient unused temp variable here) */
609 previous=get_next_block(bd);
611 bd->writeCount=previous;
612 return (previous!=RETVAL_LAST_BLOCK) ? previous : gotcount;
616 current=bd->writeCurrent;
617 goto decode_next_byte;
620 /* Allocate the structure, read file header. If in_fd==-1, inbuf must contain
621 a complete bunzip file (len bytes long). If in_fd!=-1, inbuf and len are
622 ignored, and data is read from file handle into temporary buffer. */
624 static int start_bunzip(bunzip_data **bdp, int in_fd, unsigned char *inbuf,
629 const unsigned int BZh0=(((unsigned int)'B')<<24)+(((unsigned int)'Z')<<16)
630 +(((unsigned int)'h')<<8)+(unsigned int)'0';
632 /* Figure out how much data to allocate */
634 i=sizeof(bunzip_data);
635 if(in_fd!=-1) i+=IOBUF_SIZE;
637 /* Allocate bunzip_data. Most fields initialize to zero. */
641 /* Setup input buffer */
643 if(-1==(bd->in_fd=in_fd)) {
646 } else bd->inbuf=(unsigned char *)(bd+1);
648 /* Init the CRC32 table (big endian) */
650 bd->crc32Table = crc32_filltable(1);
652 /* Setup for I/O error handling via longjmp */
654 i=setjmp(bd->jmpbuf);
657 /* Ensure that file starts with "BZh['1'-'9']." */
660 if (((unsigned int)(i-BZh0-1)) >= 9) return RETVAL_NOT_BZIP_DATA;
662 /* Fourth byte (ascii '1'-'9'), indicates block size in units of 100k of
663 uncompressed data. Allocate intermediate buffer for block. */
665 bd->dbufSize=100000*(i-BZh0);
667 bd->dbuf=xmalloc(bd->dbufSize * sizeof(int));
671 /* Example usage: decompress src_fd to dst_fd. (Stops at end of bzip data,
674 USE_DESKTOP(long long) int
675 uncompressStream(int src_fd, int dst_fd)
677 USE_DESKTOP(long long total_written = 0;)
682 outbuf=xmalloc(IOBUF_SIZE);
683 i=start_bunzip(&bd,src_fd,0,0);
686 if((i=read_bunzip(bd,outbuf,IOBUF_SIZE)) <= 0) break;
687 if(i!=write(dst_fd,outbuf,i)) {
688 i=RETVAL_UNEXPECTED_OUTPUT_EOF;
691 USE_DESKTOP(total_written += i;)
695 /* Check CRC and release memory */
697 if(i==RETVAL_LAST_BLOCK) {
698 if (bd->headerCRC!=bd->totalCRC) {
699 bb_error_msg("data integrity error when decompressing");
703 } else if (i==RETVAL_UNEXPECTED_OUTPUT_EOF) {
704 bb_error_msg("compressed file ends unexpectedly");
706 bb_error_msg("decompression failed");
712 return i ? i : USE_DESKTOP(total_written) + 0;
717 static char * const bunzip_errors[]={NULL,"Bad file checksum","Not bzip data",
718 "Unexpected input EOF","Unexpected output EOF","Data error",
719 "Out of memory","Obsolete (pre 0.9.5) bzip format not supported."};
721 /* Dumb little test thing, decompress stdin to stdout */
722 int main(int argc, char *argv[])
724 int i=uncompressStream(0,1);
727 if(i<0) fprintf(stderr,"%s\n", bunzip_errors[-i]);
728 else if(read(0,&c,1)) fprintf(stderr,"Trailing garbage ignored\n");