.NET Core Plugins

This project provides API for loading .NET Core assemblies dynamically, executing them as extensions to the main application, and finding and isolating the dependencies of the plugin from the main application. This library supports .NET Core 2, but works best in .NET Core 3 and up. It allows fine-grained control over assembly isolation and type sharing. Read more details about type sharing below.

Blog post introducing this project, July 25, 2018: .NET Core Plugins: Introducing an API for loading .dll files (and their dependencies) as 'plugins'.

Since 2018, .NET Core 3 was released, and it added stdlib API to improve assembly loading. If you are interested in understanding that API, see "Create a .NET Core application with plugins" on docs.microsoft.com. The result of this tutorial would be a simpler version of DotNetCorePlugins, but missing some features like an API for unifying types across the load context boundary, hot reload, and .NET Core 2 support.

Getting started

Install the McMaster.NETCore.Plugins NuGet package.

dotnet add package McMaster.NETCore.Plugins

The main API to use is PluginLoader.CreateFromAssemblyFile.

PluginLoader.CreateFromAssemblyFile(
    assemblyFile: "./plugins/MyPlugin/MyPlugin1.dll",
    sharedTypes: new [] { typeof(IPlugin), typeof(IServiceCollection), typeof(ILogger) },
    isUnloadable: true)

assemblyFile = the file path to the main .dll of the plugin
sharedTypes = a list of types which the loader should ensure are unified. (See What is a shared type?)
isUnloadable = (.NET Core 3+ only). Allow this plugin to be unloaded from memory at some point in the future. (Requires ensuring that you have cleaned up all usages of types from the plugin before unloading actually happens.)

See example projects in samples/ for more detailed, example usage.

Usage

Using plugins requires at least two projects: (1) the 'host' app which loads plugins and (2) the plugin, but typically also uses a third, (3) an contracts project which defines the interaction between the plugin and the host.

For a fully functional sample of this, see samples/hello-world/ .

The plugin contract

You can define your own plugin contract. A minimal contract might look like this.

public interface IPlugin
{
    string GetName();
}

There is nothing special about the name "IPlugin" or the fact that it's an interface. This is just here to illustrate a concept. Look at samples/ for additional examples of ways you could define the interaction between host and plugins.

The plugins

Typically, it is best to implement plugins by targeting net5.0 or higher. They can target netstandard2.0 as well, but using net5.0 is better because it reduces the number of redundant System.* assemblies in the plugin output.

A minimal implementation of the plugin could be as simple as this.

internal class MyPlugin1 : IPlugin
{
    public string GetName() => "My plugin v1";
}

As mentioned above, this is just an example. This library doesn't require the use of "IPlugin" or interfaces or "GetName()" methods. This code is only here to demonstrates how you can decouple hosts and plugins, but still use interfaces for type-safe interactions.

The host

The host application can load plugins using the PluginLoader API. The host app needs to define a way to find the assemblies for the plugin on disk. One way to do this is to follow a convention, such as:

plugins/
    $PluginName1/
        $PluginName1.dll
        (additional plugin files)
    $PluginName2/
        $PluginName2.dll

It is important that each plugin is published into a separate directory. This will avoid contention between plugins and duplicate dependency issues.

You can prepare the sample plugin above by running

dotnet publish MyPlugin1.csproj --output plugins/MyPlugin1/

An implementation of a host which finds and loads this plugin might look like this. This sample uses reflection to find all types in plugins which implement IPlugin, and then initializes the types' parameter-less constructors. This is just one way to implement a host. More examples of how to use plugins can be found in samples/.

using McMaster.NETCore.Plugins;

var loaders = new List<PluginLoader>();

// create plugin loaders
var pluginsDir = Path.Combine(AppContext.BaseDirectory, "plugins");
foreach (var dir in Directory.GetDirectories(pluginsDir))
{
    var dirName = Path.GetFileName(dir);
    var pluginDll = Path.Combine(dir, dirName + ".dll");
    if (File.Exists(pluginDll))
    {
        var loader = PluginLoader.CreateFromAssemblyFile(
            pluginDll,
            sharedTypes: new [] { typeof(IPlugin) });
        loaders.Add(loader);
    }
}

// Create an instance of plugin types
foreach (var loader in loaders)
{
    foreach (var pluginType in loader
        .LoadDefaultAssembly()
        .GetTypes()
        .Where(t => typeof(IPlugin).IsAssignableFrom(t) && !t.IsAbstract))
    {
        // This assumes the implementation of IPlugin has a parameterless constructor
        IPlugin plugin = (IPlugin)Activator.CreateInstance(pluginType);

        Console.WriteLine($"Created plugin instance '{plugin.GetName()}'.");
    }
}

What is a shared type?

By default, each instance of PluginLoader represents a unique collection of assemblies loaded into memory. This can make it difficult to use the plugin if you want to pass information from plugin to the host and vice versa. Shared types allow you define the kinds of objects that will be passed between plugin and host.

For example, let's say you have a simple host app like samples/hello-world/, and two plugins which were compiled with a reference interface IPlugin. This interface comes from Contracts.dll. When the application runs, by default, each plugin and the host will have their own version of Contracts.dll which .NET Core will keep isolated.

The problem with this isolation is that an object of IPlugin created within the "PluginApple" or "PluginBanana" context does not appear to be an instance of IPlugin in any of the other plugin contexts.

Configuring a shared type of IPlugin allows the .NET to pass objects of this type across the plugin isolation boundary. It does this by ignoring the version of Contracts.dll in each plugin folder, and sharing the version that comes with the Host.

Read even more details about shared types here.

Support for MVC and Razor

A common usage for plugins is to load class libraries that contain MVC controllers or Razor Pages. You can set up an ASP.NET Core to load controllers and views from a plugin using the McMaster.NETCore.Plugins.Mvc package.

dotnet add package McMaster.NETCore.Plugins.Mvc

The main API to use is .AddPluginFromAssemblyFile(), which can be chained onto the call to .AddMvc() or .AddRazorPages() in the Startup.ConfigureServices method.

public class Startup
{
    public void ConfigureServices(IServiceCollection services)
    {
        var pluginFile = Path.Combine(AppContext.BaseDirectory, "plugins/MyRazorPlugin/MyRazorPlugin.dll");
        services
            .AddMvc()
            // The AddPluginFromAssemblyFile method comes from McMaster.NETCore.Plugins.Mvc
            .AddPluginFromAssemblyFile(pluginFile);
    }
}

See example projects in samples/aspnetcore-mvc/ for more detailed, example usage.

Reflection

Sometimes you may want to use a plugin along with reflection APIs such as Type.GetType(string typeName) or Assembly.Load(string assemblyString). Depending on where these APIs are used, they might fail to load the assemblies in your plugin. In .NET Core 3+, there is an API which you can use to set the ambient context which .NET's reflection APIs will use to load the correct assemblies from your plugin.

Example:

var loader = PluginLoader.CreateFromAssemblyFile("./plugins/MyPlugin/MyPlugin1.dll");

using (loader.EnterContextualReflection())
{
    var myPluginType = Type.GetType("MyPlugin.PluginClass");
    var myPluginAssembly = Assembly.Load("MyPlugin1");
}

Read this post written by .NET Core engineers for even more details on contextual reflection.

Overriding the Default Load Context

Under the hood, DotNetCorePlugins is using a .NET Core API called ApplicationLoadContext. This creates a scope for resolving assemblies. By default, PluginLoader will create a new context and fallback to a default context if it cannot find an assembly or if type sharing is enabled. The default fallback context is inferred when PluginLoader is instantiated. In certain advanced scenarios, you may need to manually change the default context, for instance, plugins which then load more plugins, or when running .NET in a custom native host.

To override the default assembly load context, set PluginConfig.DefaultContext. Example:

AssemblyLoadContext myCustomDefaultContext = // (something).
PluginLoader.CreateFromAssemblyFile(dllPath,
     config => config.DefaultContext = myCustomDefaultContext);

wangjun009xx/DotNetCorePlugins