Exploring the Compound File Binary Format
Although the march of progress steadily tramples the old tried and true in favor of enlightened file formats designed for the new era of the web, some of us take joy in digging deeper into bits and bytes of binary file formats. I’m one of those and I can’t resist hacking my way through one of the oldest formats used by Microsoft applications, the Compound File Binary Format (CFBF).
Besides having been the bread and butter for the Microsoft Office suite of applications for many years (Visio .vsd, Publisher .pub, Outlook .msg have not replaced CFBF files as their formats in the latest versions) and almost all OLE applications that are capable of linking and embedding, CFBF has been put to use in many other applications and environments as well. Picture It!, Digital Image, SQL Server 2000 DTS packages, MSI packages, the Explorer Shell thumbnail cache, Internet Explorer RSS feeds, Rights Management add-on for Internet Explorer as well as many third party application formats like AAF (a professional file interchange format designed for video post production and authoring), Kodak FlashPix photo file, CAD programs like Solid Works just to name a few. And you can also find the reverse engineered implementations like Open Office designed for interoperability with Microsoft Office. A pure Java implementation of the OLE 2 Compound Document named POIFS (part of the Apache POI Java API To Access Microsoft Format Files), POLE and HPSF are also reverse engineered implementations. These are just a few examples the aftereffect of which should be a sinking … er, I mean warm feeling that CFBF is kind of ubiquitous.
This is why we publish [MS-CFB] Compound File Binary File Format Specification, an open specification for CFBF. By reading this entire document, you will thoroughly understand and be able to parse your way manually through any compound file using this format. Well, that’s kind of optimistic. Theoretically that is true. The document is very complete and rescues us from the days of a cryptic, hard to find document (available in the Win32 SDK long ago) to a much happier place. However, there is still some complexity inherent in the underlying structure. Assuming you have the programmer’s perception of the compound file already burned into the neurons of your brain, for example the Root Storage, Storages, Streams, names like \001CompObj, etc…, figure 1 below, gives the high level overview (like International Space Station’s view) of the organization that’s manipulated by the structured storage API’s like StgCreateDofile() and IStorage::CreateStream() on your behalf:
Figure 1: High level view of compound file
Because the details are in [MS-CFB], I’ll give the nickel tour here. FAT’s (File Allocation Table) in CFBF are pretty much like FAT’s in the file system on disk. The whole file is divided conveniently in to 512 byte (or 4096 byte, depending on the version of the format your application uses) sectors. Therefore, a FAT array maps out all the 512 byte sectors in the file. The very first sector in the file consists of a header record with a signature of 0xd0cf11ea1b11ae1 at offset 0, just so you know it’s really a compound document. Get it, docfile … d0cf11e… ok. This header points to everything else in the file. The DIFAT array, the first 109 elements of which are in the header, is the meta-FAT for FAT sectors. Because you can have more than one sector containing FAT array data (i.e. sector map), the DIFAT array tells which sectors in the file are FAT’s. The Directory Entry array simply provides a directory of all the storages and streams in the file including the “Root Entry”, a special directory entrywhich every compound file has and which serves as the parent of all other directory entries. Each stream directory entry points to the first sector containing data and the number of sectors total used by that stream. The mini stream is a special stream, pointed to by the Root Entry instead of the Directory Entry array, that contains application stream data that is less than the mini stream cutoff size (0x1000) in total length. These are mapped by a miniFAT array and each “mini sector” if you will is actually only 64 bytes in length. This saves room on fragmentation when an application’s stream only uses a fraction of the 512 or 4096 byte sectors normally allocated for data.
In the process of manually parsing files from various applications looking for non-conformance to the specification or possible logic bugs, it’s expedient to use binary editors, especially ones that will support parsing scripts. Doing it by hand is really tedious. In the way Netmon supports scripts for parsing packet records on the wire, scripts can be written to parse the structures that are described in MS-CFB. These parsing scripts can be very powerful and make life much easier. However, complexity in traversing non-contiguous sectors of a stream in script logic can cause some to gravitate toward the pragmatism of just preprocessing the file to rewrite streams with contiguous sectors. This may seem the lazy approach to script writing but it actually aids in visual verification of the stream data as well. For example, you’ve found the starting sector of a stream, let’s say “STM1”, of interest in an application’s data file. It took you very little effort to get there, you just 1) parsed the Directory Entry array, looking for the stream’s name and 2) located the starting sector number in the stream’s directory entry and voila (which is French for add one to the starting sector number and multiply by 512), you’re there. So you start reading down the line and everything looks in order but on the 512th (zero-based) byte you don’t see what you expect. Well, that’s because the next sector in the file doesn’t belong STM1!
Figure 2: Non-contiguous sectors in a stream
Now you have to go back to the FAT array and find out what sector is the next in the chain. How inconvenient. If they were all contiguous, you could just keep reading. If only you had a tool that would rearrange the stream’s sectors to be contiguous. There are tools that will defragment and compact compound files but you may not have access to one easily. I remembered reading “Inside OLE 2” by Kraig Brockschmidt about <clearing throat> years ago from cover to cover and it seemed to me that he talked about structured storage and compound files with optimism and enthusiasm not seen in the MSDN docs. So I blew the dust off the page tops and after just a few minutes found the section on defragmentation. A very interesting statement compelled me. On page 398, Brockschmidt said that IStorage::CopyTo() would, “…remove all the dead space in the process of copying the contents of one storage to another and will order the contents of streams sequentially, as shown…” and he gave a diagram to bolster my resolve. At this point, I was excited to have a possibly simply and elegant solution to the non-contiguous stream sector problem. Just use IStorage::CopyTo() to copy the root storage from one compound document to a newly created “processed” document. The recursive algorithm for CopyTo() is pretty straightforward and goes something like this:
For each source child element, create child storage/stream in target
recursively call source->CopyTo(new target IStorage)
Set target stream size
Copy target stream contents
This algorithm leads us to expect that the resulting file would have good contiguity, so I immediately coded the small fragment in a console application. Here’s what the code looks like:
int _tmain(int argc, _TCHAR* argv)
int srclen = 0;
if ((argc < 1)||((srclen = _tcslen(argv)) > 107)) // give 107 chars total, the path plus 7 for ".sequen" that we'll add
return -1; // do failure processing here.
LPSTORAGE lpSrcStg = NULL;
// pop open the docfile and suck out the root storage
HRESULT hr = StgOpenStorage(argv, NULL, STGM_DIRECT | STGM_READ | STGM_SHARE_DENY_WRITE,
NULL, NULL, &lpSrcStg);
return hr; // do failure processing here.
// This part creates the new file name, taking “file.ext” and making it “file.sequen.ext”
_TCHAR* lptszTail = _tcsrchr((_TCHAR*)argv, '.' ); // find the last dot
// allocate the new path buffer and clear it
_TCHAR* lptszTarget = new _TCHAR; // allocate enough to hold the path plus the ".sequen" insertion and a null.
memset(lptszTarget, 0x00, 108); // clear out the remainder of the string.
// copy the path up the the extension, insert ".sequen", then copy the extension.
int posdot = lptszTail-argv;
_tcsncpy_s(lptszTarget, posdot+1, argv, posdot); // copy everything before the dot.
_tcsncpy_s(lptszTarget+posdot, 8, _T(".sequen"), 7); // insert the ".sequen"
_tcscpy_s(lptszTarget+posdot+7, _tcslen(lptszTail)+1, lptszTail); // copy the rest
LPSTORAGE lpDestStg = NULL;
hr = StgCreateDocfile(lptszTarget, STGM_TRANSACTED | STGM_READWRITE | STGM_SHARE_EXCLUSIVE, NULL, &lpDestStg);
return hr; // do failure processing here.
hr = lpSrcStg->CopyTo(NULL, NULL, NULL, lpDestStg);
return hr; // do failure processing here.
Figure 3: Code to copy a compound document, reordering the stream sectors
The active code is in black. It’s very simple, just open the source file and copy the root storage to the destination. StgOpenStorage() and StgCreateDocFile() open and create compound files, respectively, returning an IStorage pointer representing the root storage to the file. According to Brockschmidt and the reference for IStorage::CopyTo() this should have some positive effect and indeed, upon testing, it does. Verifying by hand, some of the streams are actually rewritten contiguously in the process of copying . Hm… only some? Further investigation showed me that the streams that were non-contiguous still were the “small” streams. Remember the ones that I mentioned earlier that live in the ministream and are allocated with the miniFAT array? Well, these will be copied by the CopyTo() algorithm just like the larger streams. That means that from the logical perspective of the stream being written, each 64 byte sector is written sequentially until all the sectors for that stream are down on disk. However, the ministream is not just one sector and as it grows, new sectors are allocated for it. New sectors that are not necessarily contiguous. Note in figure 4 below that the sectors for STM3 are indeed written sequentially but not contiguously. Therefore, any given small stream in the ministream is only contiguous if all of it’s 64 byte sectors reside within one of the ministream’s 512 byte sectors. This means that at most a small stream can have eight 64-byte sectors of data in order to fit contiguously in one ministream sector (assuming 512 byte sectors).
Figure 4: Non-contiguous sectors of the ministream break up the (sequential) sectors of contained streams
The code shown in figure 3 will have the desired effect of organizing the sectors for larger streams both sequentially and contiguously, making them easier to read. Although it will write smaller streams (< sector size) sequentially as well, it will not guarantee that the smaller stream sectors are contiguous. That is a problem to solve in a later session.