Creating Extensible Applications with MAF (System.AddIn)

Last month, my colleague, Pinku Surana,
wrote an article about .NET AppDomains and how they can be used to provide
component isolation and make your applications more reliable.  If you missed it, you can read the Developments
archives
on Developmentor’s website.   This month, I’d like to continue exploring reliability
and extensibility by introducing you to a new framework included with .NET 3.5:
The Managed Add-in Framework (MAF), sometimes referred to as System.AddIn.

Developers (and managers) have long desired a way to easily
create extensible applications that allow new features to be added without
jeopardizing the stability of the existing code base.  The .NET framework has provided the underlying support to
accomplish this from the very beginning through the reflection API and
AppDomain support Pinku examined last month.  MAF builds on that fundamental support to provide a
higher-level service that allow you to dynamically discover, load, secure and interact
with external assemblies used to provide features for your application.  Several common architectural
requirements are made trivial with MAF:

1)  
Isolate aspects of your code for security
reasons or partial-trust scenarios.

2)  
Allow business partners to extend your
application safely without access to the source code; Adobe Illustrator is a
great example of this style of application.

3)  
Separate volatile sections of your application
out – where depending on the customer the application needs to execute
different sets of code.

4)  
Add or change code without unloading the
application – for example pay-to-play scenarios, or where you need to
update an assembly but the application must continue to run.

5)  
Develop and evolve different sections of the
application in parallel without any fear of destabilizing one based on the
other.

If any of these scenarios sound like something you could
utilize in your application then you should know about MAF!

At the center of the MAF-based application is the pipeline.

The Pipeline

Communication between add-ins and the host application is
strictly regulated by the pipeline. 
The pipeline is dynamically created by MAF through a set of
loosely-coupled components which are used to version, marshal and transform the
data as it passes back and forth between the host and each loaded add-in.

image001.png

Each section of the pipeline is contained in a separate
assembly, loaded as necessary to manage the specific add-in.  MAF discovers each component and then
loads them on request using reflection. 
 

For reliability, MAF allows optional isolation boundaries to
be created between the add-in and host – everything to the left of the
contract is loaded into the main (host) AppDomain and everything to the right
loaded into a newly created AppDomain which has its own security permission set.  The isolation can also be done through
cross-process calls if true process-level isolation is necessary.  Under the covers, the system uses
traditional remoting calls to do the work of marshaling calls back and forth.

Breaking down the pipeline, there are three main parts
starting in the center with the contract.

Contract

As you might expect, MAF is based on interface
contracts.  Interfaces allow classes
in an application to be loosely coupled, reducing the risk when changes are
made between dependent sections of the code.   The interface contract is shared between the host and
add-in and once it has been established, it should never be changed. 

Consider the simple example of a translator program.  The host will expect to load one or
more translator add-ins that will do some work on an inbound string and return
the results.  To accomplish this, I
might create an ITranslator
interface that looks like:

[AddInContract]

public interface ITranslator : IContract

{

  
string Translate(string input);   
}

Notice that the interface derives from IContract.
  This is a requirement for
add-in contracts and is what will be used to provide marshalling support when
the pipeline is established.  We
also need to decorate the interface with the [AddInContract] attribute – this is the marker used by MAF to
identify the contract when it is dynamically constructing the pipeline.  Both of these types come from the System.AddIn.Pipeline namespace in System.AddIn.Contract.dll.

Views

Moving to the edges, we find the views.  This is the code that the add-in and
host directly interact with.  It
represents a host or add-in specific “view” of the contract and, like the
contract, is contained in a separate assembly.

Both of the view classes will echo the structure of the
contract, but not actually be dependent upon the contract.  For example, our host side view might
look like this:

public abstract class TranslatorHostView

{

  
public abstract string Translate(string input);
}

The view is commonly exposed as an abstract class to make it look more natural
when used by the host, but an interface will work just as well.  The host uses the view directly when
communicating with any add-in based on the contract.

On the add-in side, we have an almost identical class
– except we decorate this type with the [AddInBase] attribute so MAF knows which side of the pipeline this
view is for (the add-in).   This
is located in the System.AddIn.Pipeline
namespace in the System.AddIn
assembly.

When we create each add-in, the implementation will use this
type as the base class.

[AddInBase]

public abstract class TranslatorAddInView

{

  
public abstract string Translate(stringinp);
}

Adapters

The last piece of the pipeline is the adapters.  The adapters play a very special role
in the pipeline – they are the glue that binds the contract to the view:
implementing lifetime management and any necessary data conversion between the
two ends. 

It may seem redundant to have this class, but by separating
the view from the contract we introduce version tolerance into our architecture
– the host can version independently of the add-in and vice-versa.  Depending on the situation, the
adapters can be as simple as a pass-through class, or can provide higher-level
services to allow non-serializable types to be marshaled across the isolation
boundaries.

On the host side, the adapter will implement the view
(remember it is either an interface or abstract class).  It will be passed a reference to the
contract in the constructor and it is responsible for connecting the two together.

[HostAdapter]

public class TranslatorHostViewToContract :  

                    
TranslatorHostView

{

  
ITranslator _contract;

  
ContractHandle _lifetime;

 

  
public TranslatorHostViewToContract(

             
ITranslator contract)  

  
{

  
   _contract =
contract;

      _lifetime = new ContractHandle(contract);

  
}

 

  
public override string Translate(string inp) 

  
{

  
   return
_contract.Translate(inp);
   }
}

In this simple example, the code caches off the contract interface.  This is a remoting proxy to the actual
loaded add-in.  The adapter then
implements the Translate method and
passes it forward to the contract for implementation.  If we had non-serializable types, then the adapter would be
responsible for converting them into something that was serializable. 

It also provides some lifetime management for the
contract.  Because the contract is
a remoting proxy and is likely running in a separate AppDomain (or even
process) we have to be concerned with how long it lives.  MAF provides all the support for this through
the ContractHandle class.  Most of the time, all you will need to
do is store a ContractHandle in your
host adapter and then pass it the inbound contract to wrap in your constructor.

Finally, in order for MAF to identify this class, it must be
decorated with the [HostAdapter]
attribute from the System.AddIn.Pipeline
namespace out of the System.AddIn.dll.

The add-in side looks very similar, but does the opposite: it
makes the add-in view look like the contract.

[AddInAdapter]

public class TranslatorAddInViewToContract :

                 
ContractBase, ITranslator

{

  
TranslatorAddInView _view;

 

  
public TranslatorAddInViewToContract(

            
TranslatorView view)  

  
{

  
   _view = view;

  
}

 

  
public string Translate(string inp)  { 

      return _view.Translate(inp); 

  
}

}

 
MAF passes the view to the constructor and the class caches the reference off
in a field.  It implements the contract
(ITranslator) and ContractBase which provides the
implementation of the IContract interface
for us (remember this was a required interface on our contract).  As the host makes calls to the contract
interface, this class will translate those calls to the add-in view, which as
you will see is the implementation provided by the add-in itself.  Note how the [AddInAdapter] attribute is used to mark this class so MAF can
discover it.

If it seems like all the above is a lot of repetitious,
boilerplate code well.. you’re right! 
To make it easier on the developer to create the pipeline components,
the MAF team has created a pipeline
builder
available at http://www.codeplex.com/clraddins.  It takes the contract assembly and then
generates the views and adapters from it:

image003.jpg

Creating an add-in

Once the pipeline pieces are built you can create
add-ins.  Each add-in provides an
implementation of the abstract add-in view.  For example, we might provide a BabelFish add-in for
universalized translation:

[AddIn(“BabelFishTranslator”,

       Description=“Universal
translator”,

 
     Version=“1.0.0.0”,
Publisher=“Zaphod Beeblebrox”]

public class BabelFishAddIn :
TranslatorAddInView

{

  
public string Translate(string input)

  
{

    

  
}
}

The add-in implements the AddInView,
providing a concrete implementation of the Translate
method.  It is decorated with the [AddIn] attribute that allows it to
provide a name, version, description and other data the host can use to decide
whether it is a useful translator.

Putting the pieces together – the directory structure

To properly identify each of the components necessary, MAF
enforces a particular directory structure you need to follow for
deployment.  Each piece is stored
in a sub-directory off the root of the pipeline directory (this is typically
the APPBASE where the host executable is stored). 

image006.png

The
directory names are required, but not case sensitive – each directory
holding a single piece of the pipeline that MAF will load dynamically when it
is loading the add-in.  The
host-side view is always located in the same directory as the host executable
itself so it does not need a dedicated directory.  When creating your Visual Studio project, it is important to
set the output directories appropriately so that you create the above directory
structure.  In addition, all
references between the components should be marked as CopyLocal = “false” in
Visual Studio to ensure a local copy of the assembly isn’t placed into the
sub-directory: 

image008.png

Discovering and loading add-ins from the host

The last piece of the puzzle is the actual loading of
add-ins from the host.  This is
done in three basic steps:

1.    
Identify and catalog the add-ins

2.    
Retrieve the list of specific add-ins based on
view or name

3.    
Activate and use the add-in

The first step is to identify the available add-ins for the
host.  This is done through the System.AddIn.AddInStore class:

string[] errorList = AddInStore.Rebuild(

  
      PipelineStoreLocation.ApplicationBase);

 

Calling Rebuild
forces MAF to walk the directory structure and create the pipeline
database.  It stores this
information in the pipeline root directory and returns a list of errors, if any
occur.  The most common errors are
missing pipeline components – where MAF is unable to locate some required
portion of the pipeline such as a View or Adapter.

Next, the host will retrieve a list of add-ins based on the
host view through the FindAddIns
method – these will be the add-ins conforming to a specific contract
(whatever the view/adapter is bound to):

Collection<AddInToken> addinTokens =
AddInStore.FindAddIns(

               
typeof(TranslatorHostView),

               
PipelineStoreLocation.ApplicationBase);

 

The first parameter is the host view type – so MAF
knows what add-ins we are looking for, the second is the pipeline root
directory, which is the same directory passed to the Rebuild method and indicates where the pipeline database is stored.

 

The returning collection represents a series of tokens that are used to identify and
control the add-ins.  This is how
the host can examine, activate and unload the add-in.  To activate a specific add-in, you can take the token and
call Activate:

 

Collection<AddInToken>
addinTokens;

   

foreach
(AddInToken token in addinTokens)
{

   TranslatorHostView view =

      
token.Activate<TranslatorHostView>(

            
AddInSecurityLevel.Internet);

  

   string hello =
view.Translate(“Bonjour”);

    
}

 

The Activate
call loads all the required components, instantiates the add-in type and returns
the host view used to interact with the add-in.  Calls made to this object will be marshaled back and forth
to the add-in using the pipeline.

 

Notice that the parameter passed to Activate indicates the security level required.  There are several overrides that allow
you to dictate exactly how the add-in is loaded.  You can load the add-in into a specific AppDomain –
the current one for best performance:

 

token.Activate<TranslatorHostView>(AppDomain.CurrentDomain)

 

You can specify a specific permission set to restrict the
things the add-in can do on your behalf:

PermissionSet
pset = …;

token.Activate<TranslatorHostView>(pset);

Or you can specify a known permission set based on the CAS zones:

image010.png

Once the add-in is activated, the host can call it just like
any other object – but never forget that you are likely crossing an
isolation barrier!  Make sure to design
your contracts so that you minimize the number of calls between the host and
add-ins to ensure your performance doesn’t suffer.

When you are finished with the add-in, you can tell MAF to
unload it through the AddInController
associated with the view:

AddInController ctrl =    

      AddInController.GetAddInController(view);

ctrl.Shutdown();

This will unload the add-in side pipeline and then destroy the containing
AppDomain so you release the resources associated with it.

There are many other capabilities MAF provides such as
versioning, passing collections and WPF visuals, passing non-serializable
types, etc.  You can get a pretty
decent overview of the features from MSDN and the
MAF team has a blog at http://blogs.msdn.com/clraddins/
that has some great, practical information on using this new framework.

About Mark Smith

Windows systems developer with low-level operating system, threading and .NET experience. For the past several years I have been involved with WPF/Silverlight and most recently iOS and Android.

I have worked as an architect and designer for several companies and am particularly interested in client/server and graphical solutions.

Specialties:.NET, WPF, Silverlight, Windows, iOS