How to handle dependency injection in a WPF/MVVM application

I am starting a new desktop application and I want to build it using MVVM and WPF.

I am also intending to use TDD.

The problem is that I don´t know how I should use an IoC container to inject my dependencies on my production code.

Suppose I have the folowing class and interface:

public interface IStorage
{
    bool SaveFile(string content);
}

public class Storage : IStorage
{
    public bool SaveFile(string content){
        // Saves the file using StreamWriter
    }
}

And then I have another class that has IStorage as a dependency, suppose also that this class is a ViewModel or a business class...

public class SomeViewModel
{
    private IStorage _storage;

    public SomeViewModel(IStorage storage){
        _storage = storage;
    }
}

With this I can easily write unit tests to ensure that they are working properly, using mocks and etc.

The problem is when it comes to use it in the real application. I know that I must have an IoC container that links a default implementation for the IStorage interface, but how would I do that?

For example, how would it be if I had the following xaml:

<Window 
    ... xmlns definitions ...
>
   <Window.DataContext>
        <local:SomeViewModel />
   </Window.DataContext>
</Window>

How can I correctly 'tell' WPF to inject dependencies in that case?

Also, suppose I need an instance of SomeViewModel from my C# code, how should I do it?

I feel I am completely lost in this, I would appreciate any example or guidance of how is the best way to handle it.

I am familiar with StructureMap, but I am not an expert. Also, if there is a better/easier/out-of-the-box framework, please let me know.

I have been using Ninject, and found that it's a pleasure to work with. Everything is set up in code, the syntax is fairly straightforward and it has a good documentation (and plenty of answers on SO).

So basically it goes like this:

Create the view model, and take the IStorage interface as constructor parameter:

class UserControlViewModel
{
	public UserControlViewModel(IStorage storage)
	{

	}
}

Create a ViewModelLocator with a get property for the view model, which loads the view model from Ninject:

class ViewModelLocator
{
	public UserControlViewModel UserControlViewModel
	{
		get { return IocKernel.Get<UserControlViewModel>();} // Loading UserControlViewModel will automatically load the binding for IStorage
	}
}

Make the ViewModelLocator an application wide resource in App.xaml:

<Application ...>
    <Application.Resources>
		<local:ViewModelLocator x:Key="ViewModelLocator"/>
	</Application.Resources>
</Application>

Bind the DataContext of the UserControl to the corresponding property in the ViewModelLocator.

<UserControl ...
			 DataContext="{Binding UserControlViewModel, Source={StaticResource ViewModelLocator}}">
    <Grid>
	</Grid>
</UserControl>

Create a class inheriting NinjectModule, which will set up the necessary bindings (IStorage and the viewmodel):

class IocConfiguration : NinjectModule
{
	public override void Load()
	{
		Bind<IStorage>().To<Storage>().InSingletonScope(); // Reuse same storage every time

		Bind<UserControlViewModel>().ToSelf().InTransientScope(); // Create new instance every time
	}
}

Initialize the IoC kernel on application startup with the necessary Ninject modules (the one above for now):

public partial class App : Application
{		
	protected override void OnStartup(StartupEventArgs e)
	{
		IocKernel.Initialize(new IocConfiguration());

		base.OnStartup(e);
	}
}

I have used a static IocKernel class to hold the application wide instance of the IoC kernel, so I can easily access it when needed:

public static class IocKernel
{
	private static StandardKernel _kernel;

	public static T Get<T>()
	{
		return _kernel.Get<T>();
	}

	public static void Initialize(params INinjectModule[] modules)
	{
		if (_kernel == null)
		{
		    _kernel = new StandardKernel(modules);
		}
	}
}

This solution does make use of a static ServiceLocator (the IocKernel), which is generally regarded as an anti-pattern, because it hides the class' dependencies. However it is very difficult to avoid some sort of manual service lookup for UI classes, since they must have a parameterless constructor, and you cannot control the instantiation anyway, so you cannot inject the VM. At least this way allows you to test the VM in isolation, which is where all the business logic is.

If anyone has a better way, please do share.

EDIT: Lucky Likey provided an answer to get rid of the static service locator, by letting Ninject instantiate UI classes. The details of the answer can be seen here

In your question you set the value of the DataContext property of the view in XAML. This requires that your view-model has a default constructor. However, as you have noted, this does not work well with dependency injection where you want to inject dependencies in the constructor.

So you cannot set the DataContext property in XAML and also do dependency injection. Instead you have other alternatives.

If you application is based on a simple hierarchical view-model you can construct the entire view-model hierarchy when the application starts (you will have to remove the StartupUri property from the App.xaml file):

public partial class App {

  protected override void OnStartup(StartupEventArgs e) {
    base.OnStartup(e);
    var container = CreateContainer();
    var viewModel = container.Resolve<RootViewModel>();
    var window = new MainWindow { DataContext = viewModel };
    window.Show();
  }

}

This is based around an object graph of view-models rooted at the RootViewModel but you can inject some view-model factories into parent view-models allowing them to create new child view-models so the object graph does not have to be fixed. This also hopefully answers your question suppose I need an instance of SomeViewModel from my cs code, how should I do it?

class ParentViewModel {

  public ParentViewModel(ChildViewModelFactory childViewModelFactory) {
    _childViewModelFactory = childViewModelFactory;
  }

  public void AddChild() {
    Children.Add(_childViewModelFactory.Create());
  }

  ObservableCollection<ChildViewModel> Children { get; private set; }

 }

class ChildViewModelFactory {

  public ChildViewModelFactory(/* ChildViewModel dependencies */) {
    // Store dependencies.
  }

  public ChildViewModel Create() {
    return new ChildViewModel(/* Use stored dependencies */);
  }

}

If your application is more dynamic in nature and perhaps is based around navigation you will have to hook into the code that performs the navigation. Each time you navigate to a new view you need to create a view-model (from the DI container), the view itself and set the DataContext of the view to the view-model. You can do this view first where you pick a view-model based on a view or you can do it view-model first where the view-model determines which view to use. A MVVM framework provides this key functionality with some way for you to hook your DI container into the creation of view-models but you can also implement it yourself. I am a bit vague here because depending on your needs this functionality may become quite complex. This is one of the core functions you get from a MVVM framework but rolling your own in a simple application will give you a good understanding what MVVM frameworks provide under the hood.

By not being able to declare the DataContext in XAML you lose some design-time support. If your view-model contains some data it will appear during design-time which can be very useful. Fortunately, you can use design-time attributes also in WPF. One way to do this is to add the following attributes to the <Window> element or <UserControl> in XAML:

xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"
xmlns:d="http://schemas.microsoft.com/expression/blend/2008"
mc:Ignorable="d"
d:DataContext="{d:DesignInstance Type=local:MyViewModel, IsDesignTimeCreatable=True}"

The view-model type should have two constructors, the default for design-time data and another for dependency injection:

class MyViewModel : INotifyPropertyChanged {

  public MyViewModel() {
    // Create some design-time data.
  }

  public MyViewModel(/* Dependencies */) {
    // Store dependencies.
  }

}

By doing this you can use dependency injection and retain good design-time support.

What I'm posting here is an improvement to sondergard's Answer, because what I'm going to tell doesn't fit into a Comment :)

In Fact I am introducing a neat solution, which avoids the need of a ServiceLocator and a wrapper for the StandardKernel-Instance, which in sondergard's Solution is called IocContainer. Why? As mentioned, those are anti-patterns.

Making the StandardKernel available everywhere

The Key to Ninject's magic is the StandardKernel-Instance which is needed to use the .Get<T>()-Method.

Alternatively to sondergard's IocContainer you can create the StandardKernel inside the App-Class.

Just remove StartUpUri from your App.xaml

<Application x:Class="Namespace.App"
             xmlns="http://schemas.microsoft.com/winfx/2006/xaml/presentation"
             xmlns:x="http://schemas.microsoft.com/winfx/2006/xaml">
             ... 
</Application>

This is the App's CodeBehind inside App.xaml.cs

public partial class App
{
    private IKernel _iocKernel;

    protected override void OnStartup(StartupEventArgs e)
    {
        base.OnStartup(e);

        _iocKernel = new StandardKernel();
        _iocKernel.Load(new YourModule());

        Current.MainWindow = _iocKernel.Get<MainWindow>();
        Current.MainWindow.Show();
    }
}

From now on, Ninject is alive and ready to fight :)

Injecting your DataContext

As Ninject is alive, you can perform all kinds of injections, e.g Property Setter Injection or the most common one Constructor Injection.

This is how you inject your ViewModel into your Window's DataContext

public partial class MainWindow : Window
{
    public MainWindow(MainWindowViewModel vm)
    {
        DataContext = vm;
        InitializeComponent();
    }
}

Of course you can also Inject an IViewModel if you do the right bindings, but that is not a part of this answer.

Accessing the Kernel directly

If you need to call Methods on the Kernel directly (e.g. .Get<T>()-Method), you can let the Kernel inject itself.

    private void DoStuffWithKernel(IKernel kernel)
    {
        kernel.Get<Something>();
        kernel.Whatever();
    }

If you would need a local instance of the Kernel you could inject it as Property.

    [Inject]
    public IKernel Kernel { private get; set; }

Allthough this can be pretty useful, I would not recommend you to do so. Just note that objects injected this way, will not be available inside the Constructor, because it's injected later.

According to this link you should use the factory-Extension instead of injecting the IKernel (DI Container).

> The recommended approach to employing a DI container in a software system is that the Composition Root of the application be the single place where the container is touched directly.

How the Ninject.Extensions.Factory is to be used can also be red here.

I go for a "view first" approach, where I pass the view-model to the view's constructor (in its code-behind), which gets assigned to the data context, e.g.

public class SomeView
{
    public SomeView(SomeViewModel viewModel)
    {
        InitializeComponent();

        DataContext = viewModel;
    }
}

This replaces your XAML-based approach.

I use the Prism framework to handle navigation - when some code requests a particular view be displayed (by "navigating" to it), Prism will resolve that view (internally, using the app's DI framework); the DI framework will in turn resolve any dependencies that the view has (the view model in my example), then resolves its dependencies, and so on.

Choice of DI framework is pretty much irrelevant as they all do essentially the same thing, i.e. you register an interface (or a type) along with the concrete type that you want the framework to instantiate when it finds a dependency on that interface. For the record I use Castle Windsor.

Prism navigation takes some getting used to but is pretty good once you get your head around it, allowing you to compose your application using different views. E.g. you might create a Prism "region" on your main window, then using Prism navigation you would switch from one view to another within this region, e.g. as the user selects menu items or whatever.

Alternatively take a look at one of the MVVM frameworks such as MVVM Light. I've got no experience of these so can't comment on what they're like to use.

Install MVVM Light.

Part of the installation is to create a view model locator. This is a class which exposes your viewmodels as properties. The getter of these properties can then be returned instances from your IOC engine. Fortunately, MVVM light also includes the SimpleIOC framework, but you can wire in others if you like.

With simple IOC you register an implementation against a type...

SimpleIOC.Default.Register<MyViewModel>(()=> new MyViewModel(new ServiceProvider()), true);

In this example, your view model is created and passed a service provider object as per its constructor.

You then create a property which returns an instance from IOC.

public MyViewModel
{
    get { return SimpleIOC.Default.GetInstance<MyViewModel>; }
}

The clever part is that the view model locator is then created in app.xaml or equivalent as a data source.

<local:ViewModelLocator x:key="Vml" />

You can now bind to its 'MyViewModel' property to get your viewmodel with an injected service.

Hope that helps. Apologies for any code inaccuracies, coded from memory on an iPad.

Canonic DryIoc case

Answering an old post, but doing this with DryIoc and doing what I think is a good use of DI and interfaces (minimal use of concrete classes).

1. The starting point of a WPF app is App.xaml, and there we tell what is the inital view to use; we do that with code behind instead of the default xaml:

2. remove StartupUri="MainWindow.xaml" in App.xaml

3. in codebehind (App.xaml.cs) add this override OnStartup:

   protected override void OnStartup(StartupEventArgs e)
   {
       base.OnStartup(e);
       DryContainer.Resolve<MainWindow>().Show();
   }
   

that's the startup point; that's also the only place where resolve should be called.

4. the configuration root (according to Mark Seeman's book Dependency injection in .NET; the only place where concrete classes should be mentionned) will be in the same codebehind, in the constructor:

   public Container DryContainer { get; private set; }
   
   public App()
   {
       DryContainer = new Container(rules => rules.WithoutThrowOnRegisteringDisposableTransient());
       DryContainer.Register<IDatabaseManager, DatabaseManager>();
       DryContainer.Register<IJConfigReader, JConfigReader>();
       DryContainer.Register<IMainWindowViewModel, MainWindowViewModel>(
           Made.Of(() => new MainWindowViewModel(Arg.Of<IDatabaseManager>(), Arg.Of<IJConfigReader>())));
       DryContainer.Register<MainWindow>();
   }
   

Remarks and few more details

• I used concrete class only with the the view MainWindow;
• I had to specify which contructor to use (we need to do that with DryIoc) for the ViewModel, because the default constructor needs to exist for the XAML designer, and the constructor with injection is the actual one used for the application.

The ViewModel constructor with DI:

public MainWindowViewModel(IDatabaseManager dbmgr, IJConfigReader jconfigReader)
{
    _dbMgr = dbmgr;
    _jconfigReader = jconfigReader;
}

ViewModel default constructor for design:

public MainWindowViewModel()
{
}

The codebehind of the view:

public partial class MainWindow
{
	public MainWindow(IMainWindowViewModel vm)
	{
		InitializeComponent();
		ViewModel = vm;
	}

	public IViewModel ViewModel
	{
		get { return (IViewModel)DataContext; }
		set { DataContext = value; }
	}
}

and what is needed in the view (MainWindow.xaml) to get a design instance with ViewModel:

d:DataContext="{d:DesignInstance local:MainWindowViewModel, IsDesignTimeCreatable=True}"

Conclusion

We hence got a very clean and minimal implementation of a WPF application with a DryIoc container and DI while keeping design instances of views and viewmodels possible.

Use the Managed Extensibility Framework.

[Export(typeof(IViewModel)]
public class SomeViewModel : IViewModel
{
    private IStorage _storage;
    
    [ImportingConstructor]
    public SomeViewModel(IStorage storage){
        _storage = storage;
    }

    public bool ProperlyInitialized { get { return _storage != null; } }
}

[Export(typeof(IStorage)]
public class Storage : IStorage
{
    public bool SaveFile(string content){
        // Saves the file using StreamWriter
    }
}

//Somewhere in your application bootstrapping...
public GetViewModel() {
     //Search all assemblies in the same directory where our dll/exe is
     string currentPath = Path.GetDirectoryName(Assembly.GetExecutingAssembly().Location);
     var catalog = new DirectoryCatalog(currentPath);
     var container = new CompositionContainer(catalog);
     var viewModel = container.GetExport<IViewModel>();
     //Assert that MEF did as advertised
     Debug.Assert(viewModel is SomViewModel); 
     Debug.Assert(viewModel.ProperlyInitialized);
}

In general, what you would do is have a static class and use the Factory Pattern to provide you with a global container (cached, natch).

As for how to inject the view models, you inject them the same way you inject everything else. Create an importing constructor (or put a import statement on a property/field) in the code-behind of the XAML file, and tell it to import the view model. Then bind your Window's DataContext to that property. Your root objects you actually pull out of the container yourself are usually composed Window objects. Just add interfaces to the window classes, and export them, then grab from the catalog as above (in App.xaml.cs... that's the WPF bootstrap file).

I would suggest to use the ViewModel - First approach https://github.com/Caliburn-Micro/Caliburn.Micro

see: https://caliburnmicro.codeplex.com/wikipage?title=All%20About%20Conventions

use Castle Windsor as IOC container.

All About Conventions

One of the main features of Caliburn.Micro is manifest in its ability to remove the need for boiler plate code by acting on a series of conventions. Some people love conventions and some hate them. That’s why CM’s conventions are fully customizable and can even be turned off completely if not desired. If you are going to use conventions, and since they are ON by default, it’s good to know what those conventions are and how they work. That’s the subject of this article. View Resolution (ViewModel-First)

Basics

The first convention you are likely to encounter when using CM is related to view resolution. This convention affects any ViewModel-First areas of your application. In ViewModel-First, we have an existing ViewModel that we need to render to the screen. To do this, CM uses a simple naming pattern to find a UserControl1 that it should bind to the ViewModel and display. So, what is that pattern? Let’s just take a look at ViewLocator.LocateForModelType to find out:

public static Func<Type, DependencyObject, object, UIElement> LocateForModelType = (modelType, displayLocation, context) =>{
    var viewTypeName = modelType.FullName.Replace("Model", string.Empty);
    if(context != null)
    {
        viewTypeName = viewTypeName.Remove(viewTypeName.Length - 4, 4);
        viewTypeName = viewTypeName + "." + context;
    }

    var viewType = (from assmebly in AssemblySource.Instance
                    from type in assmebly.GetExportedTypes()
                    where type.FullName == viewTypeName
                    select type).FirstOrDefault();

    return viewType == null
        ? new TextBlock { Text = string.Format("{0} not found.", viewTypeName) }
        : GetOrCreateViewType(viewType);
};

Let’s ignore the “context” variable at first. To derive the view, we make an assumption that you are using the text “ViewModel” in the naming of your VMs, so we just change that to “View” everywhere that we find it by removing the word “Model”. This has the effect of changing both type names and namespaces. So ViewModels.CustomerViewModel would become Views.CustomerView. Or if you are organizing your application by feature: CustomerManagement.CustomerViewModel becomes CustomerManagement.CustomerView. Hopefully, that’s pretty straight forward. Once we have the name, we then search for types with that name. We search any assembly you have exposed to CM as searchable via AssemblySource.Instance.2 If we find the type, we create an instance (or get one from the IoC container if it’s registered) and return it to the caller. If we don’t find the type, we generate a view with an appropriate “not found” message.

Now, back to that “context” value. This is how CM supports multiple Views over the same ViewModel. If a context (typically a string or an enum) is provided, we do a further transformation of the name, based on that value. This transformation effectively assumes you have a folder (namespace) for the different views by removing the word “View” from the end and appending the context instead. So, given a context of “Master” our ViewModels.CustomerViewModel would become Views.Customer.Master.

Remove the startup uri from your app.xaml.

App.xaml.cs

public partial class App
{
    protected override void OnStartup(StartupEventArgs e)
    {
        IoC.Configure(true);
     
        StartupUri = new Uri("Views/MainWindowView.xaml", UriKind.Relative);

        base.OnStartup(e);
    }
}

Now you can use your IoC class to construct the instances.

MainWindowView.xaml.cs

public partial class MainWindowView
{
    public MainWindowView()
    {
        var mainWindowViewModel = IoC.GetInstance<IMainWindowViewModel>();

        //Do other configuration            
        
        DataContext = mainWindowViewModel;

        InitializeComponent();
    }

}

Another simple solution is to create a murkup extension that resolves your view model by its type:

public class DISource : MarkupExtension {
    public static Func<Type, object, string, object> Resolver { get; set; }

    public Type Type { get; set; }
    public object Key { get; set; }
    public string Name { get; set; }

    public override object ProvideValue(IServiceProvider serviceProvider) => Resolver?.Invoke(Type, Key, Name);
}

You can adjust this extension to any DI container in the following manner:

protected override void OnStartup(StartupEventArgs e) {
    base.OnStartup(e);
    DISource.Resolver = Resolve;
}
object Resolve(Type type, object key, string name) {
    if(type == null)
        return null;
    if(key != null)
        return Container.ResolveKeyed(key, type);
    if(name != null)
        return Container.ResolveNamed(name, type);
    return Container.Resolve(type);
}

Use it in XAML as simple as that:

DataContext="{local:DISource Type=viewModels:MainViewModel}"

This way, you will be able to easily assign DataContext to your view and automatically inject all required parameters directly to your view model using your DI container. With this technique, you don't have to pass your DI container or other parameters to the View constructor.

DISource doesn't depend on a container type, so you can use it with any Dependency Injection Framework. It's sufficient to set the DISource.Resolver property to a method that knows how to use your DI container.

I described this technique in greater detail at Dependency Injection in a WPF MVVM Application