Microsoft's HTML Help (.chm) format

Preface

This is documentation on the .chm format used by Microsoft HTML Help. This format has been reverse engineered in the past, but as far as I know this is the first freely available documentation on it. One Usenet message indicates that these .chm files are actually IStorage files documented in the Microsoft Platform SDK. However, I have not been able to locate such documentation.

Note

The word "section" is badly overloaded in this document. Sorry about that.

All numbers are in hexadecimal unless otherwise indicated in the text. Except in tabular listings, this will be indicated by $ or 0x as appropriate. All values within the file are Intel byte order (little endian) unless indicated otherwise.

The overall format of a .chm file

The .chm file begins with a short ($38 byte) initial header. This is followed by the header section table and the offset to the content. Collectively, this is the "header".

The header is followed by the header sections. There are two header sections. One header section is the file directory, the other contains the file length and some unknown data. Immediately following the header sections is the content.

The header starts with the initial header, which has the following format

0000: char[4]  'ITSF'
0004: DWORD    3 (Version number)
0008: DWORD    Total header length, including header section table and
               following data.
000C: DWORD    1 (unknown)
0010: DWORD    a timestamp.
               Considered as a big-endian DWORD, it appears to contain
               seconds (MSB) and fractional seconds (second byte).
	       The third and fourth bytes may contain even more fractional
               bits.  The 4 least significant bits in the last byte are
               constant.
0014: DWORD    Windows Language ID.  The two I've seen
               $0409 = LANG_ENGLISH/SUBLANG_ENGLISH_US
               $0407 = LANG_GERMAN/SUBLANG_GERMAN
0018: GUID     {7C01FD10-7BAA-11D0-9E0C-00A0-C922-E6EC}
0028: GUID     {7C01FD11-7BAA-11D0-9E0C-00A0-C922-E6EC}

Note: a GUID is $10 bytes, arranged as 1 DWORD, 2 WORDs, and 8 BYTEs.

It is followed by the header section table, which is 2 entries, where each entry is $10 bytes long and has this format:

0000: QWORD    Offset of section from beginning of file
0008: QWORD    Length of section

Following the header section table is 8 bytes of additional header data. In Version 2 files, this data is not there and the content section starts immediately after the directory.

0000: QWORD    Offset within file of content section 0

The Header Sections

Header Section 0

This section contains the total size of the file, and not much else

0000: DWORD    $01FE (unknown)
0004: DWORD    0 (unknown)
0008: QWORD    File Size
0010: DWORD    0 (unknown)
0014: DWORD    0 (unknown)

Header Section 1: The Directory Listing

The central part of the .chm file: A directory of the files and information it contains.

Directory header

The directory starts with a header; its format is as follows:

0000: char[4]  'ITSP'
0004: DWORD    Version number 1
0008: DWORD    Length of the directory header
000C: DWORD    $0a (unknown)
0010: DWORD    $1000    Directory chunk size
0014: DWORD    "Density" of quickref section, usually 2.
0018: DWORD    Depth of the index tree
               1 there is no index, 2 if there is one level of PMGI
	       chunks.
001C: DWORD    Chunk number of root index chunk, -1 if there is none
               (though at least one file has 0 despite there being no
	       index chunk, probably a bug.) 
0020: DWORD    Chunk number of first PMGL (listing) chunk
0024: DWORD    Chunk number of last PMGL (listing) chunk
0028: DWORD    -1 (unknown)
002C: DWORD    Number of directory chunks (total)
0030: DWORD    Windows language ID
0034: GUID     {5D02926A-212E-11D0-9DF9-00A0C922E6EC}
0044: DWORD    $54 (This is the length again)
0048: DWORD    -1 (unknown)
004C: DWORD    -1 (unknown)
0050: DWORD    -1 (unknown)

The Listing Chunks

The header is directly followed by the directory chunks. There are two types of directory chunks -- index chunks, and listing chunks. The index chunk will be omitted if there is only one listing chunk. A listing chunk has the following format:

0000: char[4]  'PMGL'
0004: DWORD    Length of free space and/or quickref area at end of
               directory chunk 
0008: DWORD    Always 0. 
000C: DWORD    Chunk number of previous listing chunk when reading
               directory in sequence (-1 if this is the first listing chunk)
0010: DWORD    Chunk number of next listing chunk when reading
               directory in sequence (-1 if this is the last listing chunk)
0014: Directory listing entries (to quickref area)  Sorted by
      filename; the sort is case-insensitive.

The quickref area is written backwards from the end of the chunk. One quickref entry exists for every n entries in the file, where n is calculated as 1 + (1 << quickref density). So for density = 2, n = 5.

Chunklen-0002: WORD     Number of entries in the chunk
Chunklen-0004: WORD     Offset of entry n from entry 0
Chunklen-0006: WORD     Offset of entry 2n from entry 0
Chunklen-0008: WORD     Offset of entry 3n from entry 0
...

The format of a directory listing entry is as follows

      ENCINT: length of name
      BYTEs: name  (UTF-8 encoded)
      ENCINT: content section
      ENCINT: offset
      ENCINT: length

The offset is from the beginning of the content section the file is in, after the section has been decompressed (if appropriate). The length also refers to length of the file in the section after decompression.

There are two kinds of file represented in the directory: user data and format related files. The files which are format-related have names which begin with '::', the user data files have names which begin with "/".

The Index Chunk

An index chunk has the following format

0000: char[4]  'PMGI'
0004: DWORD    Length of quickref/free area at end of directory chunk
0008: Directory index entries (to quickref/free area)

The quickref area in an PMGI is the same as in an PMGL

The format of a directory index entry is as follows

      ENCINT: length of name
      BYTEs: name  (UTF-8 encoded)
      ENCINT: directory listing chunk which starts with name

When higher-level indexes exist (when the depth of the index tree is 3 or higher), presumably the upper-level indexes will contain the numbers of lower-level index chunks rather than listing chunks

Encoded Integers

An ENCINT is a variable-length integer. The high bit of each byte indicates "continued to the next byte". Bytes are stored most significant to least significant. So, for example, $EA $15 is (((0xEA&0x7F)<<7)|0x15) = 0x3515.

The Content

In Version 3, the content typically immediately follows the header sections, and is at the location indicated by the DWORD following the header section table. In Version 2, the content immediately follows the header. All content section 0 locations in the directory are relative to that point. The other content sections are stored WITHIN content section 0.

The Namelist file

There exists in content section 0 and in the directory a file called "::DataSpace/NameList". This file contains the names of all the content sections. The format is as follows:

0000: WORD     Length of file, in words
0002: WORD     Number of entries in file

Each entry:
0000: WORD     Length of name in words, excluding terminating null
0002: WORD     Double-byte characters
xxxx: WORD     0

Yes, the names have a length word AND are null terminated; sort of a belt-and-suspenders approach. The coding system is likely UTF-16 (little endian).

The section names seen so far are

"Uncompressed" is self-explanatory. The section "MSCompressed" is compressed with Microsoft's LZX algorithm.

The Section Data

For each section other than 0, there exists a file called '::DataSpace/Storage/<Section Name>/Content'. This file contains the compressed data for the section. So, conceptually, getting a file from a nonzero section is a multi-step process. First you must get the content file from section 0. Then you decompress (if appropriate) the section. Then you get the desired file from your decompressed section.

Other section format-related files

There are several other files associated with the sections

Appendix: The Compression

The compressed sections are compressed using LZX, a compression method Microsoft also uses for its cabinet files. To ensure this, check the second DWORD of compression info in the ControlData file for the section — it should be 'LZXC'. To decompress, first read the file "::DataSpace/Storage/<SectionName>/Transform/{7FC28940-9D31-11D0-9B27-00A0C91E9C7C}/InstanceData/ResetTable". This reset table has the following format

0000: DWORD    2     unknown (possibly a version number)
0004: DWORD    Number of entries in reset table
0008: DWORD    8     Size of table entry (bytes)
000C: DWORD    $28   Length of table header (area before table entries)
0010: QWORD    Uncompressed Length
0018: QWORD    Compressed Length
0020: QWORD    0x8000 block size for locations below
0028: QWORD    0 (zeroth entry of table)
0030: QWORD    location in compressed data of 1st block boundary in
               uncompressed data

Repeat to end of file

Now you can finally obtain the section (from its Content file). The window size for the LZX compression is 16 (decimal) on all the files seen so far. This is specified by the DWORD at $10 in the ControlData file (but note that DWORD gives the window size in 0x8000-byte blocks, not the LZX code for the window size)

The rule that the input bit-stream is to be re-aligned to a 16-bit boundary after $8000 output characters have been processed IS in effect, despite this LZX not being part of a CAB file. The reset table tells you when this was done, though there is no need for that during decompression; you can just keep track of the number of output characters. Furthermore, while this does not appear to be documented in the LZX format, the uncompressed stream is padded to an $8000 byte boundary.

There is one change from LZX as defined by Microsoft: After each LZX reset interval (defined in the ControlData file, but in practice equal to the window size) of compressed data is processed, the LZX state is fully reset, as if an entirely new file was being encoded. This allows semi-random access to the compressed data; you can start reading on any reset interval boundary using the reset interval size and the reset table.

Note:
Earlier versions of this document stated that the reset interval only reset the Huffman tables and required outputting the 1-bit header again. This was erroneous. The Lempel Ziv state is reset as well. In practice, a decoder works just as well with the incorrect assumption, but encoding a file with match positions which refer to a time before the most recent LZX reset causes crashes on decoding.


Acknowledgements

The following people in (no particular order) have submitted information which has helped correct and close the gaps in this document.

And others I have not been able to reach.


Copyright 2001-2003 Matthew T. Russotto

You may freely copy and distribute unmodified copies of this file, or copies where the only modification is a change in line endings, padding after the html end tag, coding system, or any combination thereof. The original is in ASCII with Unix line endings.