Dr. Dobb's is part of the Informa Tech Division of Informa PLC

This site is operated by a business or businesses owned by Informa PLC and all copyright resides with them. Informa PLC's registered office is 5 Howick Place, London SW1P 1WG. Registered in England and Wales. Number 8860726.


Channels ▼
RSS

.NET

Launcher: Mastering Your Own Domain


October, 2004: Launcher: Mastering Your Own Domain

Loading and unloading managed programs into multiple domains

Gigi is a software developer specializing in object- and component-oriented programming using C++. He can be contacted at gigi_sayfanplaystation.sony.com.


Programmer productivity hinges on many factors, including talent, problem-domain experience, programming language, and tools, quality of tools, and—most importantly—the level of focus and degree of concentration, which is in constant flux. The ideal state of mind to program in is in the "flow," where you are fully focused on the problem, time flies, and you perform amazing programming feats. Some people call this "being in the zone" or "deep hack mode." Unfortunately, the smallest interruption can break the flow. If you work with the wrong tools or programming language, you might never achieve flow—you will instead fight your development environment, try to figure out obscure compiler messages, browse help files, and spend mental effort on side issues.

For instance, a friend and I are developing an online game using C# for the .NET platform. There is a server and multiple clients that share interfaces and common data types through a shared assembly. My development cycle includes editing the code, building the solution, and launching the server/client programs for testing. However, it takes time to launch a process, load the CLR, load the necessary system assemblies, load the server/client assembly, and start running. This ceremony gets old really fast, interrupts my flow, and strikes me as unnecessary. I should only load/unload the modified code.

The CLR provides the means to do this via AppDomains. A single OS process can host multiple AppDomains. The novel thing is that not only can DLLs be loaded into an AppDomain, but executable programs, too. Consequently, a single process can host multiple managed programs. In this article, I present Launcher, an application that loads/unloads other managed programs into multiple AppDomains in its own process, executes them, and unloads them on demand. This application solves my problem because I can load/unload only my client and server assemblies into existing processes. Moreover, common system assemblies may already be loaded into the process, so the server and client programs don't have to load them again when they start executing. In addition to faster load/unload time, the memory footprint is reduced because the CLR and many assemblies that had to be loaded for each program in its own process can now be shared. While this discussion applies to CLR 1.1, I also mention some possible changes in Whidbey.

AppDomains, Processes, and Threads

The CLR is a managed runtime inside a managed runtime, which is the hosting OS. The OS has its own execution and isolation model for running code, which uses processes and OS threads. The CLR has a separate execution, security, and isolation model, which uses AppDomains and CLR threads. AppDomains are much lighter then OS processes. When a managed program starts executing, it is usually loaded into the default AppDomain. However, the code in the default AppDomain may create other AppDomains in the same process, and load and execute other programs into them. The CLR verifies managed code upon just-in-time (JIT) compilation to be typesafe and guarantees that a fault in one AppDomain doesn't affect other ones in the same process.

A managed program runs inside a regular Win32 process that hosts the CLR as a COM object, loads the managed program, and executes it. This means that the CLR is subjected to the rules and regulations of the OS. Specifically, code always executes in the context of some OS thread. The CLR has its own concept of threads that are separate from OS threads. The mapping between OS threads and CLR threads is not documented officially, and it is considered an implementation detail that may change in a future version of the CLR. CLR threads may cross AppDomains, but they have a separate thread local storage (TLS) in each AppDomain (accessible through the Thread.GetData and Thread.SetData APIs). The TLS of the OS thread associated with the CLR thread has its own AppDomain-agnostic TLS, which never changes when crossing AppDomains. In the current implementation (Version 1.1), the same CLR thread is associated with the same OS thread whenever the OS thread enters the same AppDomain, but don't count on it.

When a program (executable assembly) is loaded and executed in an AppDomain, it usually loads many other assemblies. By default, each AppDomain maintains its own copy in memory of the loaded assemblies even if other AppDomains in the same process have already loaded these assemblies. The only assembly that is always shared between all AppDomains is mscorlib.

The loader optimization attribute loads assemblies to be shared between AppDomains. The benefit is a smaller working set and quicker load time if many assemblies are shared. The downside is that access to static fields is slightly slower for shared assemblies because an indirection is needed to find the static field address of the active AppDomain. Regardless of the sharing mode, every AppDomain always maintains a distinct copy of static fields to facilitate the isolation of AppDomains. If static fields were shared, then code in one AppDomain that modified a shared static field might wreak havoc in another AppDomain, which might not suspect it. Another disadvantage of shared assemblies is that they cannot be unloaded.

The loader optimization attribute tells the JIT compiler how assemblies should be loaded. There are three values:

  • SingleDomain (the default) causes the AppDomain to load a private copy of each necessary assembly's code.
  • MultiDomain loads one copy of each assembly per process.
  • MultiDomainHost loads all strongly named assemblies per-process, although in future versions, it may only load assemblies from the global assembly cache (GAC) per-process.

I recommend the MultiDomainHost option for most multidomain scenarios. Private assemblies that are used by a single program should not be shared (since the performance of static fields will suffer) at the process level. MultiDomainHost is reasonable since GAC assemblies are usually used by many programs. Because most applications use only the default AppDomain, it makes sense that the default is SingleDomain to get fast access to static fields.

Working with AppDomains

The AppDomain class exposes several methods, properties, and events for creating AppDomains, getting information on their state, unloading AppDomains, and communicating between AppDomains. Here, I show how to create a new AppDomain, load a program into it, and execute it. Then I handle events, exchange data between AppDomains, execute code, and create objects in another AppDomain.

The first step is to create a strongly named assembly that can be shared in a process by multiple AppDomains. I created a dummy key using the strong-name tool (sn.exe), added a postbuild event to the project of the shared assembly to install it to the GAC for easy resolution, and changed a couple of attributes in AssemblyInfo.cs; see Listing One and the comments in AssemblyInfo.cs for details.

Because I want my multidomain program to share assemblies, I set the loader optimization attribute to MultiDomainHost (Listing Two). Shared assemblies must be strongly named. A strongly named assembly has a cryptographic signature that is useful in security-oriented code scenarios.

Next, I create a new AppDomain by calling the AppDomain.CreateDomain static method. This method has several overloaded versions, the simplest one just accepts a domain name. There are many attributes that control the behavior of AppDomain, which can be specified through the AppDomainSetup class. Once an AppDomain has been created, it can also be destroyed or unloaded. The terminology is a little inconsistent, but it hints to asymmetry in the way code and resources are managed in the CLR. You can load specific assemblies into an AppDomain, but you cannot unload an assembly from an AppDomain. The only way to unload an assembly is to unload the entire AppDomain (with all the other assemblies that are loaded into it). Listing Three shows how to create an AppDomain and immediately unload it.

But just creating new AppDomains and unloading them isn't very interesting—I want to run some code in the new AppDomain. There are many ways to load code into the AppDomain and execute it. The simplest method is to call the ExecuteAssembly method of the new AppDomain (Listing Four). This method loads the specified assembly and executes it immediately in the new AppDomain. The calling thread is blocked until the program in the new AppDomain exits. It is possible to use asynchronous method invocation to allow the calling thread to continue working. The DomainMaster library (available electronically; see "Resource Center," page 5) implements this approach.

Okay, you can launch programs in new AppDomains, but what if you want information on their execution? The AppDomain class exposes several events (see Table 1). You can register for these events just like any other CLR event. The important thing is, of course, to register the event handlers before loading the program because most events occur before the program actually starts executing (Listing Five). The event handlers execute in the AppDomain they were registered in. This means that the code of the event handlers must be loaded into that AppDomain. The assembly loader must be able to locate the assembly that contains the event handlers. This is an involved process, but the bottom line is that the assembly must be either on the probe path (which is always under the application base directory), under the dynamic base directory (if specified), or in the GAC.

Events are cool but they are handled in the context of the AppDomain that they occur in. You may want to communicate information between AppDomains and there are various ways to do so. One category is through OS-level IPC (simple files, P/Invoke Win32 IPC); the other category is AppDomain-specific communication methods, which is what I focus on here. The simplest method is using the GetData/SetData methods. You can Get/Set a value or a full-fledged object from/to a different AppDomain. Objects are passed by value (a copy of the original object) or by reference (a proxy to the original object). Listing Six demonstrates passing the creating AppDomain object to the created AppDomain. The GetData/SetData methods are intuitive to use and provide a dictionary API where each value/object is identified by a string.

The CLR allows executing arbitrary code in a different AppDomain via the DoCallback() method. The code requires defining a delegate of type CrossAppDomainDelegate. Listing Seven demonstrates an interaction between two AppDomains. The default AppDomain creates another AppDomain and calls SetData() to pass itself to the other AppDomain. Then it executes some code in the other AppDomain that uses the data that was set earlier. The assembly that contains the CrossAppDomainDelegate code, as well as any other assemblies it references, must be loaded into the other AppDomain prior to calling the DoCallback() method.

The most advanced way of communication is creating instances of objects in a different domain. When creating an instance using the CreateInstance() method, the return value is an ObjectHandle. which can't be used directly in the current AppDomain without unwrapping it first. The benefit of an ObjectHandle is that it does not require loading all the metadata of the wrapped object. It decouples the creation of the object from its use. The creating AppDomain A may blindly create an object instance O in AppDomain B, then pass O to AppDomain C where it is unwrapped and used. The metadata of object O does not have to be loaded into AppDomain A. A simpler scenario is when AppDomain A creates an object instance O in AppDomain B and would like to use it. Listing Eight demonstrates that by using the CreateInstanceAndUnwrap() method to create an AppDomainExplorer instance in the other AppDomain. Since AppDomainExplorer is derived from MarshalByRefObject, the returned object is a proxy to an instance in the other AppDomain and not a local copy. Creating AppDomain loads the assembly, which contains the type and its metadata, into the other AppDomain.

What actually happens when an AppDomain is unloaded? The DomainUnload event is called and all threads are aborted. However, due to the subtle interactions between OS threads and CLR threads, some threads cannot be aborted. In this case, an exception is thrown that claims that a thread cannot be unwound, but there is no good way to handle it. The AppDomain won't be unloaded along with all its assemblies and that's it. I encountered this situation when I called Console.ReadLine(). When a nondefault AppDomain is unloaded, an AppDomainUnloadException is thrown in the default AppDomain as an unhandled exception. I tried to catch it explicitly by wrapping the call to AppDomain.Unload with a try-catch block, but it didn't help. I guess the exception is thrown in the unloaded AppDomain, which is already in a state that doesn't let it handle it. The text of the unhandled exception is: "An unhandled exception of type 'System.AppDomainUnloadedException' occurred in Unknown Module." GUI applications cannot be unloaded cleanly without quitting first. If you try to unload an active GUI application, it crashes unpleasantly.

In general, the Console doesn't play nicely with multidomain programs. The main reason is that there is one console per process and it is not trivial to share it properly between multiple console apps in various AppDomains. This is a basic feature of the OS—one console at most per one process. Even using Win32 APIs, you can just replace a console with another console or create a console for a GUI app (AllocConsole()), but you can't have more than one console per process. So, all the console apps in a multidomain program share the same console. If the default AppDomain contains a GUI app, then a console is created for the first console app that is loaded into another AppDomain. The title of the console always reflects the latest console application that was loaded. There is no way to switch "focus" between apps in different AppDomains. This leads to a serious problem with Console.ReadLine(). When users press the Enter key, which app in which AppDomain() should handle it? It turns out to be a moot point because, even if there is a single app that called Console.ReadLine(), pressing Enter has no effect. Threads that wait on ReadLine() cannot be aborted and their AppDomains won't unload. Finally, if the app in your default AppDomain is a GUI app and you load a console app into another AppDomain, you will not be able to get rid of the console. Even if you never call Console.ReadLine() and unload all the AppDomains that contain console apps, the console window will remain there. It sort of becomes your main window. A console window created by closing the main form doesn't terminate the program, while closing the console window does, regardless of whether there are still active AppDomains that contain console apps. I reported all these issues to Chris Brumme (a senior architect on the CLR team), who provided me with invaluable inside information and said that he is aware that Console.ReadLine() is blocked in an OS API that cannot be safely interrupted (http://blogs.msdn.com/cbrumme/). These issues are not likely to be solved in Whidbey despite the serious overhaul the Console is getting.

Launcher

Launcher (available electronically) is a .NET GUI app that loads and executes other programs in multiple app domains and allows unloading and reloading them. It is an extraordinary deployment and debugging tool. Launcher's GUI lets users add new applications to the managed list. Applications on the list can be executed (loaded) or stopped (unloaded). The application list is stored in a file so Launcher remembers the applications across invocations.

Launcher is really just a GUI front end to the DomainMaster engine, which lives in a separate assembly (DomainMasterLib). The DomainMaster engine is a Singleton. This modularization cleanly separates the core AppDomain management from the GUI front end. It also allows experimenting with different front ends or easily creating a dedicated custom Launcher. For example, always launch the BoogaBooga server, wait five seconds, launch three BoogaBooga clients, and finally launch the standalone Logger application.

The public interface of the DomainMasterLib assembly is defined in the Interfaces.cs file (available electronically). You acquire the main interface (IDomainMaster) from the static property DomainMaster.Instance. This is the only way since the constructors are private. DomainMaster stores a little information about every application in the AppInfo class: the executable path, the AppDomain object it belongs to (null if the application is not running), and a state enumeration (IDLE, RUNNING, SHUTTING_DOWN, or NONE). It exposes an outgoing interface (IDomainMasterEvents) for notifying the front end on interesting events such as assembly load and AppDomain unload. The IDomainMaster interface is the active interface, which the front end uses to communicate with DomainMaster. The front end may add new applications to the managed list of DomainMaster, it may launch idle applications, terminate running applications, and get a collection of all the managed apps through the Apps property. To receive the events, the front end needs to implement the IDomainMasterEvents interface and call the AttachSink() method.

The API is not perfect. You can't remove an application once it was added (except trough editing the persistent apps.txt file) and the fact that the DomainMaster is a Singleton allows only a single applications collection and a single sink (per process). However, it is good enough for my needs and it can be easily modified if necessary.

The DomainMaster class operates much like the sample code previously discussed. The major difference is that the DomainMaster is using asynchronous method calls to avoid blocking its users. The ExecuteAssemblyDelegate is declared at the top, and in the Launch() method I wrap it around the new AppDomain ExecuteAssembly method and invoke it asynchronously using BeginInvoke(). I don't wait or check the result of the invocation. If something went wrong, I get an exception. The rest of the code is just bookkeeping of the apps collection. The event handlers notify the user through the IDomainMasterEvents interface about interesting events and the Terminate() method is used to unload running applications.

I include two test applications called Client and Server (available electronically). These applications talk through Remoting and can be launched through Launcher. Remoting is the official way to communicate between applications in different AppDomains, which may be in the same process, a different process on the same machine, or on a different machine altogether. The test applications are both console programs and demonstrate all the console-related issues I keep whining about.

Conclusion

In the future, Windows will likely move to direct mapping of OS concepts to CLR concepts. The current discrepancies are difficult to overcome—this is clear when trying to fit AppDomains into the OS world of processes, hard threads, and consoles. This probably improves the performance because managed-unmanaged costs a lot today (about 50 CPU instructions). As more and more APIs are migrated to managed code, this becomes less of a problem because you will not have to leave the managed world most of the time. Still, for interoperability with legacy Win32 and COM code, it is important to improve this aspect of the CLR.

How about Launcher? I hope that Whidbey solves the console issues I mentioned. But, in the meantime, a couple of possible directions I intend to follow include intercepting console events and output and writing them to a GUI console simulator, and hosting the CLR as a COM object and using the low-level interfaces it provides to improve the behavior of Launcher.

DDJ



Listing One

// The post build event
call "$(DevEnvDir)..\Tools\vsvars32.bat"
sn -Vr $(TargetDir)$(TargetFileName)
gacutil -u $(TargetName)
gacutil -i $(TargetDir)$(TargetFileName)

// From AssemblyInfo.cs
[assembly: AssemblyKeyFileAttribute("..\\..\\public.snk")]
[assembly: AssemblyDelaySignAttribute(true)]
Back to article


Listing Two
namespace DomainMasterSample
{
   class MultiDomainApp
   {
      [LoaderOptimization(LoaderOptimization.MultiDomainHost)]
      static void Main() 
      {
      }
}
Back to article


Listing Three
namespace DomainMasterSample
{
   class MultiDomainApp
   {
      [LoaderOptimization(LoaderOptimization.MultiDomainHost)]
      static void Main() 
      {
         // Create a new AppDoamin
         AppDomain ad = AppDomain.CreateDomain("Playground Domain");         
         // Unload the domain
         AppDomain.Unload(ad);         
      }

   }
}
Back to article


Listing Four
namespace DomainMasterSample
{
   class MultiDomainApp
   {
      [LoaderOptimization(LoaderOptimization.MultiDomainHost)]
      static void Main() 
      {
         // Create a new AppDomain
         Console.WriteLine(AppDomain.CurrentDomain.FriendlyName);
         AppDomain ad = AppDomain.CreateDomain("Playground Domain");         
         // Register domain event handlers
         ad.AssemblyLoad += new AssemblyLoadEventHandler(AssemblyLoadHandler);
         ad.DomainUnload += new EventHandler(DomainUnloadHandler);
         // Load and execute assembly in domain         
         ad.ExecuteAssembly("Playground.exe");
         // Unload the domain
         AppDomain.Unload(ad);         
      }
      static public void AssemblyLoadHandler(object sender, 
                                                 AssemblyLoadEventArgs args) 
      {
         AppDomain ad = sender as AppDomain;
         Debug.Assert(ad == AppDomain.CurrentDomain);
         string a = args.LoadedAssembly.GetName().Name;         
         AppDomain c = ad.GetData("Creator") as AppDomain;
         Console.WriteLine("Assembly Loaded: '{0}' in '{1}' 
                     created by '{2}'", a, ad.FriendlyName, c.FriendlyName);
      }
      static public void DomainUnloadHandler(object sender, EventArgs args) 
      {
         AppDomain ad = sender as AppDomain;
         Debug.Assert(ad == AppDomain.CurrentDomain);         
      }
   }
}
Back to article


Listing Five
namespace DomainMasterSample
{
   class MultiDomainApp
   {
      [LoaderOptimization(LoaderOptimization.MultiDomainHost)]
      static void Main() 
      {
         // Create a new AppDomain
         Console.WriteLine(AppDomain.CurrentDomain.FriendlyName);
         AppDomain ad = AppDomain.CreateDomain("Playground Domain");         
         // Register domain event handlers
         ad.AssemblyLoad += new AssemblyLoadEventHandler(AssemblyLoadHandler);
         ad.DomainUnload += new EventHandler(DomainUnloadHandler);
         // Load and execute assembly in domain         
         ad.ExecuteAssembly("Playground.exe");
         // Unload the domain
         AppDomain.Unload(ad);         
      }
      static public void AssemblyLoadHandler(object sender, 
                                                  AssemblyLoadEventArgs args) 
      {
         AppDomain ad = sender as AppDomain;
         Debug.Assert(ad == AppDomain.CurrentDomain);
         string a = args.LoadedAssembly.GetName().Name;         
         AppDomain c = ad.GetData("Creator") as AppDomain;
         Console.WriteLine("Assembly Loaded: '{0}' in '{1}' 
                      created by '{2}'", a, ad.FriendlyName, c.FriendlyName);
      }
      static public void DomainUnloadHandler(object sender, EventArgs args) 
      {
         AppDomain ad = sender as AppDomain;
         Debug.Assert(ad == AppDomain.CurrentDomain);         
      }
   }
}
Back to article


Listing Six
namespace DomainMasterSample
{
   class MultiDomainApp
   {
      [LoaderOptimization(LoaderOptimization.MultiDomainHost)]
      static void Main() 
      {
         // Create a new AppDomain
         Console.WriteLine(AppDomain.CurrentDomain.FriendlyName);
         AppDomain ad = AppDomain.CreateDomain("Playground Domain");         
         // Exchange data between AppDomains
         ad.SetData("Creator", AppDomain.CurrentDomain);
         AppDomain creator = ad.GetData("Creator") as AppDomain;
         Debug.Assert(creator == AppDomain.CurrentDomain);
         // Unload the domain
         AppDomain.Unload(ad);         
      }
   }
}
Back to article


Listing Seven
namespace DomainMasterSample
{
   class MultiDomainApp
   {
      [LoaderOptimization(LoaderOptimization.MultiDomainHost)]
      static void Main() 
      {
         // Create a new AppDomain
         Console.WriteLine(AppDomain.CurrentDomain.FriendlyName);
         AppDomain ad = AppDomain.CreateDomain("Playground Domain");         
         // Exchange data between AppDomains
         ad.SetData("Creator", AppDomain.CurrentDomain);
         // Execute some code in a different AppDomain
         CrossAppDomainDelegate cadd = 
                     new CrossAppDomainDelegate(CrossAppDomainCallback);
         ad.DoCallBack(cadd);
         // Unload the domain
         AppDomain.Unload(ad);         
      }
      static public void CrossAppDomainCallback()
      {
         AppDomain ad = AppDomain.CurrentDomain;
         AppDomain c = ad.GetData("Creator") as AppDomain;
         Console.WriteLine("CrossAppDomainCallback() running in '{0}' 
                       created by '{1}'", ad.FriendlyName, c.FriendlyName);
      }
   }
}
Back to article


Listing Eight
namespace DomainMasterSample
{
   public class AppDomainExplorer : MarshalByRefObject
   {
      public void Explore()
      {
         AppDomain ad = AppDomain.CurrentDomain;
         Console.WriteLine("----- AppDomainExplorer.Explore() -----");
         Console.WriteLine("Name: {0}", ad.FriendlyName);
         Console.WriteLine("Setup Info: {0}", ad.SetupInformation.ToString());
         Console.WriteLine("Assemblies:");
         foreach (Assembly a in ad.GetAssemblies())
         {
            Console.WriteLine(a.FullName);
         }
      }
   }
   class MultiDomainApp
   {
      [LoaderOptimization(LoaderOptimization.MultiDomainHost)]
      static void Main() 
      {
         // Create a new AppDomain
         Console.WriteLine(AppDomain.CurrentDomain.FriendlyName);
         AppDomain ad = AppDomain.CreateDomain("Playground Domain");
         // Create object in another AppDomain
         string path = Path.Combine(Directory.GetCurrentDirectory(), 
                                                 "DomainMasterSample.exe");
         Assembly a = Assembly.LoadFile(path);
         ad.Load(a.GetName());
         AppDomainExplorer ade = 
                ad.CreateInstanceAndUnwrap(a.GetName().FullName, 
                "DomainMasterSample.AppDomainExplorer") as AppDomainExplorer;
         ade.Explore();
         // Unload the domain
         AppDomain.Unload(ad);
         Console.ReadLine();
      }
   }
}
Back to article


Related Reading


More Insights






Currently we allow the following HTML tags in comments:

Single tags

These tags can be used alone and don't need an ending tag.

<br> Defines a single line break

<hr> Defines a horizontal line

Matching tags

These require an ending tag - e.g. <i>italic text</i>

<a> Defines an anchor

<b> Defines bold text

<big> Defines big text

<blockquote> Defines a long quotation

<caption> Defines a table caption

<cite> Defines a citation

<code> Defines computer code text

<em> Defines emphasized text

<fieldset> Defines a border around elements in a form

<h1> This is heading 1

<h2> This is heading 2

<h3> This is heading 3

<h4> This is heading 4

<h5> This is heading 5

<h6> This is heading 6

<i> Defines italic text

<p> Defines a paragraph

<pre> Defines preformatted text

<q> Defines a short quotation

<samp> Defines sample computer code text

<small> Defines small text

<span> Defines a section in a document

<s> Defines strikethrough text

<strike> Defines strikethrough text

<strong> Defines strong text

<sub> Defines subscripted text

<sup> Defines superscripted text

<u> Defines underlined text

Dr. Dobb's encourages readers to engage in spirited, healthy debate, including taking us to task. However, Dr. Dobb's moderates all comments posted to our site, and reserves the right to modify or remove any content that it determines to be derogatory, offensive, inflammatory, vulgar, irrelevant/off-topic, racist or obvious marketing or spam. Dr. Dobb's further reserves the right to disable the profile of any commenter participating in said activities.

 
Disqus Tips To upload an avatar photo, first complete your Disqus profile. | View the list of supported HTML tags you can use to style comments. | Please read our commenting policy.