Application Isolation
When writing an application, it is often necessary to isolate parts of the application so that a failure of one part does not cause a failure in another part of the application. In Windows, application isolation has traditionally been at the process level. In other words, if a process is stopped or crashes, other processes will continue running. One process cannot directly address memory in another process's address space; however, several interprocess communication mechanisms may be utilized.
Unfortunately, it is expensive for an application to use separate processes to achieve such isolation. To switch from one process to another, the process state information (context) must be saved and restored. This includes a thread and a process switch. A thread switch requires saving CPU registers, such as the call stack and instruction pointer, and loading the information for a new thread, as well as updating the scheduling information for the threads. A process switch includes IO buffers, accounting information, and processor rights that have to be saved for the old process and restored for the new one.
Application Domain
The .NET application domain (sometimes called the AppDomain) is a more lightweight unit for application isolation, fault tolerance, and security. Multiple application domains can run in one process. Since the .NET code can be checked for type safety and security, the CLR can guarantee that one application domain can run independently of another application domain in the same process. No process switch is required to achieve application isolation.
Application domains can have multiple contexts, but a context exists in only one application domain. Although a thread runs in one context of one application domain at a time, the Threading examples Step 3 and Step 4 demonstrate that a thread can execute in more than one context. One or more threads can run in an application domain at the same time. An object lives in only one context.
Each application domain starts with a single thread and one context. Additional threads and contexts are created as needed.
There is no relationship between the number of application domains and threads. A Web server might require an application domain for each hosted application that runs in its process. The number of threads in that process would be far fewer, depending on how much actual concurrency the process can support.
To enforce application isolation, code in one application domain cannot make direct calls into the code (or even reference resources) in another application domain. They must use proxies.
Application Domains and Assemblies
Applications are built from one or more assemblies, but each assembly is loaded into an application domain. Each application domain can be unloaded independently of the others, but you cannot unload an individual assembly from an application domain. The assembly will be unloaded when the application domain is unloaded. Unloading an application domain also frees all resource associated with that application domain.
Each process has a default application domain that is created when the process is started. This default domain can only be unloaded when the process shuts down.
Applications such as ASP.NET or Internet Explorer critically depend on preventing the various applications that run under it from interfering with each other. By never loading application code into the default domain, they can ensure that a crashing program will not bring down a host.
AppDomain Class
The AppDomain class abstracts application domains. The AppDomain sample illustrates the use of application domains. This class has static methods for creating and unloading application domains:
AppDomain *domain = AppDomain::CreateDomain( "CreatedDomain2", 0, 0); ... AppDomain::Unload(domain);
While the CreateDomain method is overloaded, one signature illustrates application domain isolation:
static AppDomain CreateDomain( String *friendlyName, Evidence *securityInfo, AppDomainSetup *info);
The Evidence parameter is a collection of the security constraints on the application domain. While we will discuss this in greater detail in Chapter 13, the domain's creator can modify this collection to control the permissions that the executing application domain can have. The AppDomainSetup parameter specifies setup information about the domain. Among the information specified is the location of the application domain's configuration file and where private assemblies are loaded. Hence, application domains can be configured independently of each other. Code isolation, setup isolation, and control over security combine to ensure that application domains are independent of each other.
AppDomain Events
To help in maintaining application isolation, the AppDomain class allows you to set up event handlers for
when a domain unloads.
when the process exits.
when an unhandled exception occurs.
when attempts to resolve assemblies, types, and resources fail.
AppDomain Example
If you run the AppDomain example,15 you will get the following output in Figure 81.
Figure 81 Output of AppDomain example.
First, the name, thread, and context of the default domain is written out.
AppDomain *currentDomain = AppDomain::CurrentDomain; Console::WriteLine( "At startup, Default AppDomain is {0} ThreadId: {1} @@ ContextId {2}\n", currentDomain->FriendlyName, Thread::CurrentThread->GetHashCode().ToString(), Thread::CurrentContext->ContextID.ToString());
We then load and execute an assembly. This code in this assembly just prints out a string and its domain's name, thread, and context. Notice that it executes in the default domain.
int val = domain->ExecuteAssembly( "TestApp\\bin\\Debug\\TestApp.exe", 0, args);
We then create an instance of the Customers type from the Customer assembly in the default domain. The CreateInstance method of the AppDomain class returns an ObjectHandle instance. You can pass this ObjectHandle between application domains without loading the metadata associated with the wrapped type. When you want to use the create object instance, you must unwrap it by calling the Unwrap method on the ObjectHandle instance.
ObjectHandle *oh = currentDomain->CreateInstance( "Customer", "OI.NetCpp.Acme.Customers"); ... Customers *custs = dynamic_cast<Customers *>(oh->Unwrap());
We then add a new customer, and then list all the existing customers. Notice that both the constructor of this type and its methods execute in the same thread and context as the default domain does.
We then create a new domain and create an instance of the same type as before in that new domain.
AppDomain *domain = AppDomain::CreateDomain( "CreatedDomain1", 0, 0); ... oh = domain->CreateInstance( "Customer", "OI.NetCpp.Acme.Customers"); ... Customers *custs2 = dynamic_cast <Customers *>(oh->Unwrap());
Note that the constructor call that results from the CreateInstance method executes in the new domain and is therefore in a different context from where the CreateInstance call was made, but is executing on the same thread that made the CreateInstance call.
When we list the customers in this new object, we get a different list of customers. This is not surprising, since it is a different Customers object. Nonetheless, the customer list method executes in the default domain!
Using RemotingServices::IsTransparentProxy, we see that the ObjectHandle is a proxy to the Customers object that lives in the newly created AppDomain. However, when you unwrap the object to get an instance handle, you do not get a proxy, but you get an actual object reference. By default, objects are marshaled by value (copied) from one application domain to another.
If the Customers object is not serializable, you will get an exception when you try to copy it. This exception would be thrown when you do the Unwrap, not the CreateInstance. The latter returns a reference. The copy is made only when the ObjectHandle is unwrapped. If the object cannot be serialized, it cannot be copied from one application domain to another.
Next, we create a new thread, and that thread creates a new application domain, and then loads and executes an assembly. The assembly starts executing at its entry point, the Main routine of the AppDomainTest class.
AppDomain *domain = AppDomain::CreateDomain( "CreatedDomain2", 0, 0); ... String * args[] = new String *[1]; args[0] = "MakeReservation"; ... int val = domain->ExecuteAssembly( "TestApp\\bin\\Debug\\TestApp.exe", 0, args); ... AppDomain::Unload(domain);
The Main routine loads the Hotel assembly into the newly created application domain. In this example, the TestApp.exe application is implemented in C#. It then queries the metadata of the assembly for the HotelBroker type information. It then uses that type information to create a HotelBroker object. The HotelBroker class is marked with a synchronization attribute. As a result, the HotelBroker constructor and the MakeReservation method run in a different context than the default context.
Assembly a = AppDomain.CurrentDomain.Load("Hotel"); Type typeHotelBroker = a.GetType("OI.NetCpp.Acme.HotelBroker"); HotelBroker hotelBroker = (HotelBroker)Activator.CreateInstance(typeHotelBroker); DateTime date = DateTime.Parse("12/2/2001"); ReservationResult rr = hotelBroker.MakeReservation(1, "Boston", "Sheraton", date, 3); Console.WriteLine("\tReservation Id: {0}", rr.ReservationId);
Marshaling, AppDomains, and Contexts
By default, objects are copied from one application domain to another (marshal by value). The section "Remoting" shows how to marshal by reference between application domains. This ensures that code in one application domain is isolated from another.
Objects are marshaled by reference between contexts. This allows the CLR to enforce the requirements (such as synchronization or transactions) of different objects. This is true whether the client of the object is in the same application domain or not.
Since most objects do not derive from ContextBoundObject, they can reside or move from one context to another as required. Threads can cross application domain and context boundaries within the same Win32 process.