What's in a IFilter?
IFilter interface was designed by Microsoft for use in its Indexing Service. Its main purpose is to extract text from files so the Indexing Service can index them and later search them. Some versions of Windows comes with
IFilter implementations for Office files, and there are free and commercial filters available for other file types (Adobe PDF filter is a popular one). The
IFilter interface is used mainly in non-text files like Office documents, PDF documents etc., but is also used for text files like HTML and XML, to extract only the important parts of the file. Although the
IFilter interface can be used for general purpose text extraction from documents, it is generally used in search engines. Windows Desktop Search uses filters to index files. For more information on
IFilter, see the Links section.
So what else is new?
There are already quite a few articles and pieces of information on how to use the
IFilter interface in .NET (see the Links section), so why write another article you ask? Well, there are some problems with the implementations offered in those articles (details below) which caused me to take a different approach to using and loading filters. I'm currently using this implementation in a new product I'm developing (more details will be revealed here), and since it's working great, I decided to share it with you (yes, You!).
Issues with the current implementations
These are the issues I and others have found with the current implementations, I'll discuss each in detail below:
- Extracting text from very large files.
- COM threading issues.
- Adobe PDF filter crashing the application when it's closed.
Extracting text from very large files
All of the sample code I found on using
IFilters in C# provided a method that extracts the entire text of a document and returns that as a string. Usually, it's something like this:
public static string GetTextFromFile(string path)
Now, this might be OK for some uses, but for a general purpose indexer, I find that it isn't the most scalable way to extract text from documents. Some documents may be very large (30 MB PDFs or Word documents are not uncommon), and extracting the entire text at once can have negative effects on the garbage collector since these objects will be stored in the .NET "Large Objects Heap" (see the Links section for more information).
COM threading issues
Since filters are essentially COM objects, they carry with them all the COM threading model issues that we all love to hate. See the Links section for some of the reported problems. To make a long story short, some filters are marked as STA (Adobe PDF filter), some as MTA (Microsoft XML filter), and some as Both (Microsoft Office Filter). This means MTA filters will not load into C# threads that are marked with the
[STAThread], and STA filters will not load into
[MTAThread] threads. Some people recommend manually changing the registry to mark "problematic" filters as Both, but this isn't something you want to do during the installation of a product, nor can you reliably do it because you don't know which filters are installed on the customer's machine. We basically need a way to load an IFilter and use it no matter what its threading model or our threading model is.
Adobe PDF filter crashing the application when it's closed
There are quite a few reports about problems with the Adobe PDF filter v6. See this and this for some examples. I researched this issue for some time, and I believe I found what the problem is. It seems Adobe forgot (or not..) to export the
DllCanUnloadNow function from their PDFFILT.dll. Since a filter is implemented as a COM object, it should export this function to let COM know when it can unload this library. It seems that this causes problems for C# applications because the .dll is never unloaded, and when it does, it's probably a bit late.
In a previous version of my application, I managed to work around this issue by specifically unloading the PDFFILT.dll library. In the current implementation, this workaround is not needed.
How my implementation solves these issues?
Implement a FilterReader
I decided to go the hard way and implement a
TextReader derived class called
FilterReader. This solves issue #1 because we don't have to extract the entire text at once. Instead, you can simply use the reader to get a buffer at a time. If you still want to get the entire text as a string, use the
In order to get an
IFilter instance, you should call the
LoadIFilter API and pass it a file name.
LoadIFilter eventually calls
CoCreateInstance() to actually instantiate the filter, and thus abide to COM rules. To avoid the threading issues, I decided to bypass COM and instantiate the filter COM class myself. This has the following implications and assumptions:
- I needed to find the correct COM class that implements the filter for a specific file type.
- I needed to dynamically load the COM DLL that implements that COM class and call the
DllGetClassObject function that is exported from that .dll.
- I didn't want to re-implement all of the COM infrastructure, so in order to solve the issue of unloading COM DLLs only when they're not needed, I decided to keep the DLLs loaded during the entire application lifetime and only unload them when the application dies. Note that this essentially solves issue #3 since we manually unload the PDFFILT.dll.
IFilter should not be used by multiple threads since it is no longer protected by COM.
- I assumed that STA filters will behave correctly when called from an MTA thread when COM is not involved. Until now, I didn't encounter any problem with this approach. If you find a filter that behaves badly when used this way, please let me know.
How to use the code
Using the code is very simple: instantiate a
FilterReader by passing it the file you want to extract text from, and use it like any
TextReader derived class:
TextReader reader=new FilterReader(fileName);
Finding the correct COM class
Since I've decided not to use
LoadIFilter, I needed to find a way to locate the correct DLL and class ID of the object implementing the filter for the file whose text we're interested in. This was a simple task, thanks to the excellent RegMon utility from SysInternals. I simply called
LoadIFilter and traced which registry keys where read during that operation. I then used the same logic in my own implementation. The details can be found in the
FilterLoader class. When a class\DLL pair is found for a certain file extension, this information is cached to avoid traversing the registry again.
During the research I made on how
LoadIFilter works, I came across a utility called IFilter Explorer that shows which filters are installed on your computer. From that tool, I also learned that some indexing engines use methods not implemented in
LoadIFilter to find filters. One of these methods uses the content type registered for that extension. My version of
LoadIFilter also handles loading filters for files that have no filter registered for them but do have a filter registered for their content type.
Loading the DLL and instantiating the filter implementation
OK, so we have the name of the DLL and the ID of the class implementing our filter, how do we create an instance of that class? Most of the work is handled by the
ComHelper class. The steps needed to accomplish that are:
- Load the DLL using the
LoadLibrary Win32 API.
- Call the
GetProcAddress Win32 API to get a pointer to the
Marshal.GetDelegateForFunctionPointer() to convert that function pointer to a delegate. Note: this is only available in .NET 2.0. For an equivalent method in .NET 1.1, see the Links section.
- Call the
DllGetClassObject function to get an
private static IClassFactory GetClassFactoryFromDll(string dllName,
Guid filterPersistGUID=new Guid(filterPersistClass);
if (dllGetClassObject(ref filterPersistGUID,
ref IClassFactoryGUID, out unk)!=0)
return (unk as IClassFactory);
Once we have an
IClassFactory object, we can use it to create instances of the class implementing our filter:
private static IFilter LoadFilterFromDll(string dllName,
Guid IFilterGUID=new Guid("89BCB740-6119-101A-BCB7-00DD010655AF");
classFactory.CreateInstance(null, ref IFilterGUID, out obj);
return (obj as IFilter);
We finally have an
IFilter instance that can be passed to our
FilterReader (after doing the standard filter initialization code):
IPersistFile persistFile=(filter as IPersistFile);
IFILTER_INIT iflags =
if (filter.Init(iflags, 0, IntPtr.Zero, out flags)==IFilterReturnCode.S_OK)
Note that because we didn't use any COM calls during that process, we get a "raw" interface pointer to the filter class and COM does not create any proxy\stubs to protect that interface.
I've been using this approach for several months now without any problems. Here's a summary of the benefits and implications with this approach:
- No COM threading issues that cause certain filters not to function correctly.
- No need to mark your thread as
[STAThread] when using filters (this is a problem especially with web applications).
- The Adobe PDF filter does not crash at the end of the application.
- Better scalability when dealing with large files.
- Better filter search logic than
LoadIFilter (using content type).
- Bypassing COM may damage your health - Actually, I very much enjoy bypassing COM, but FDA regulations force me to have that warning here.
- Once filter DLLs are loaded into your application, they will stay loaded. If this is a problem for you, don't use this approach.
- No COM protection for multi-threaded access to filters (Yeah, so?).
Links and References