Object Lifetime Manager: A Complementary Pattern for Controlling Object Creation and Destruction

CREATIONAL PATTERNS SUCH as Singleton and Factory Method1 address object construction and initialization, but do not consider object destruction. However, in some applications, object destruction is as important as object construction. The Object Lifetime Manager pattern addresses issues associated with object destruction. Object Lifetime Manager (OLM) is also an example of a complementary pattern, which completes or extends other patterns. In particular, the OLM pattern completes creational patterns by considering the entire lifetime of objects.

THE OBJECT LIFETIME MANAGER PATTERN
Intent The Object Lifetime Manager pattern can be used to govern the entire lifetime of objects, from creating them prior to their first use to ensuring they are destroyed properly at program termination. In addition, this pattern can be used to replace static object creation/destruction with dynamic object preallocation/deallocation that occurs automatically during application initialization/termination.

Example Singleton1 is a common creational pattern that provides a global point of access to a unique class instance and defers creation of the instance until it is first accessed. If a singleton is not needed during the lifetime of a program, it will not be created. The Singleton pattern does not address the issue of when its instance is destroyed, however, which is problematic for certain applications and operating systems.

To illustrate why it is important to address destruction semantics, consider the following logging component that provides a client programming API to a distributed logging service.2 Applications use the logging component as a front end to the distributed logging service to report errors and generate debugging traces.

class Logger
{
public:
	// Global access point to Logger singleton.
	static Logger *instance (void) {
		if (instance_ == 0)
			instance_ = new Logger;
		return instance_;
	}

	// Write some information to the log.
	int log (const char *format, ...);

	protected:
	// Default constructor (protected to
	// ensure Singleton pattern usage).
	Logger (void);

	static Logger *instance_;
	// Contained Logger singleton instance.

	// . . . other resources that are
	//       held by the singleton . . .
};

	// Initialize the instance pointer.
Logger *Logger::instance_ = 0;
The Logger constructor, which is omitted for brevity, allocates various OS endsystem resources, such as socket handles, shared memory segments, and/or system-wide semaphores, that are used to implement the logging service client API.

To reduce the size and improve the readability of its logging records, an application may choose to log certain data, such as timing statistics, in batch mode rather than individually. For instance, the following statistics class batches timing data for individual identifiers:

class Stats
{
public:
	// Global access point to the
	// statistics singleton.
	static Stats *instance (void) {
		if (instance_ == 0)
			instance_ = new Stats;
		return instance_;
	}

	// Record a timing data point.
	int record (int id,
				const timeval &tv);

	// Report recorded statistics
	// to the log.
	void report (int id) {
		Logger::instance ()->
			log ("Avg timing %d: "
				"%ld sec %ld usec\n",
				id,
				average_i (id).tv_sec,
				average_i (id).tv_usec);
	}

	protected:
	// Default constructor.
	Stats (void);

	// Internal accessor for an average.
	const timeval &average_i (void);

	// Contained Stats singleton instance.
	static Stats *instance_;

	// . . . other resources that are
	//       held by the instance . . .
};

	// Initialize the instance pointer.
Stats *Stats::instance_ = 0;
After recording various statistics, a program calls the Stats::report method, which uses the Logger singleton to report average timing statistics for an identifier.

Both the Logger and Stats classes provide distinct services to the application: The Logger class provides general logging capabilities, whereas the Stats class provides specialized batching and logging of time statistics. These classes are designed using the Singleton pattern, so that a single instance of each is used in an application process.

The following example illustrates how an application might use the Logger and Stats singletons:

int main (int argc, char *argv[])
{
	// Interval timestamps.
	timeval start_tv, stop_tv;

	// Logger, Stats singletons
	// do not yet exist.

	// Logger and Stats singletons created
	// during the first iteration.
	for (int i = 0; i < argc;="" ++i)="" {="" ::gettimeofday="" (&start_tv);="" do="" some="" work="" between="" timestamps="" .="" .="" .="" ::gettimeofday="" (&stop_tv);="" then="" record="" the="" stats="" .="" .="" .="" timeval="" delta_tv;="" delta_tv.sec="stop_tv.sec" -="" start_tv.sec;="" delta_tv.usec="stop_tv.usec" -="" start_tv.usec;="" stats::instance="" ()-="">record (i, delta_tv);

		// . . . and log some output.
		Logger::instance ()->
			log ("Arg %d [%s]\n", i, argv[i]);
		Stats::instance()->report (i);
	}

	// Logger and Stats singletons are not
	// cleaned up when main returns.
	return 0;
}
Note that the Logger and Stats singletons are not constructed or destroyed explicitly by the application, i.e., their lifetime management is decoupled from the application logic. It is common practice to not destroy singletons at program exit.3

Several drawbacks arise, however, because the Singleton pattern only addresses the creation of singleton instances and does not deal with their destruction. In particular, when the main program above terminates, neither the Logger nor the Stats singletons are cleaned up. At best, this can lead to false reports of leaked memory. At worst, important system resources may not be released and destroyed properly.

For instance, problems can arise if the Logger and/or Stats singletons hold OS resources such as system-scope semaphores, I/O buffers, or other allocated OS resources. Failure to clean up these resources gracefully during program shutdown can cause deadlocks and other synchronization hazards. To alleviate this problem, each singleton's destructor should be called before the program exits.

One way of implementing the Singleton pattern that attempts to ensure singleton destruction is to employ the Scoped Locking C++ idiom4 that declares a static instance of the class at file scope.3 For example, the following Singleton Destroyer template provides a destructor that deletes the singleton.

template 
Singleton_Destroyer
{
public:
	Singleton_Destroyer (void): t_ (0) {}
	void register (T *) { t_ = t; }
	~Singleton_Destroyer (void) { delete t_; }
private:
	T *t_; // Holds the singleton instance.
};
To use this class, all that's necessary is to modify the Logger and Stats classes by defining a static instance of the Singleton Destroyer, such as the following example for Logger:

static Singleton_Destroyer
	logger_destroyer;

	// Global access point to the
// Logger singleton.
static Logger *instance (void) {
	if (instance_ == 0) {
		instance_ = new Logger;
		// Register the singleton so it will be
		// destroyed when the destructor of
		// logger_destroyer is run.
		logger_destroyer.register (instance_);
	}
	return instance_;
}

	// . . . similar changes to Stats class . . .
Note how logger_destroyer class holds the singleton and deletes it when the program exits. A similar Singleton_Destroyer could be used by the Stats singleton, as well.

Unfortunately, there are several problems with explicitly instantiating static Singleton_Destroyer instances. In C++, for example, each Singleton_Destroyer could be defined in a different compilation unit. In this case, there is no guaranteed order in which their destructors will be called, which can lead to undefined program behavior. In particular, if singletons in different compilation units share resources, such as socket handles, shared memory segments, and/or system-wide semaphores, the program may fail to exit cleanly. The undefined order of singleton destruction in C++ makes it hard to ensure these resources are released by the OS before 1) the last singleton using the resource is completely destroyed, but not before 2) a singleton that is still alive uses the resource(s).

In summary, the key forces that are not resolved in these examples above are 1) resources allocated by a singleton must ultimately be released when a program exits, 2) unconstrained creation and destruction order of static instances can result in serious program errors, and 3) shielding software developers from responsibility for details of object lifetime management can make systems less error-prone.

Context An application or system where full control over the lifetime of the objects it creates is necessary for correct operation.

Problem Many applications do not handle the entire lifetime of their objects properly. In particular, applications that use creational patterns, such as Singleton, often fail to address object destruction. Similarly, applications that use static objects to provide destruction often suffer from inconsistent initialization and termination behavior. These problems are outlined here.

Problems with single destruction. Singleton instances may be created dynamically.* A dynamically allocated singleton instance becomes a resource leak if it is not destroyed. Often, singleton leaks are ignored because 1) they aren't significant in many applications and 2) on most multiuser general-purpose operating systems, such as UNIX or Windows NT, they are cleaned up when a process terminates.

Unfortunately, resource leaks can be troublesome in the following contexts:

  • When graceful shutdown is required3: Singletons may be responsible for system resources such as system-wide locks, open network connections, and shared memory segments. Explicit destruction of these singletons may be desirable to ensure these resources are destroyed at a well-defined point during program termination. For instance, if the Logger class in the subsection Example requires system-wide locks or shared memory, then it should release these resources after they are no longer needed.
  • When singletons maintain references to other singletons: Explicitly managing the order of destruction of singletons may be necessary to avoid problems due to dangling references during program termination. For example, if the Stats class in the example above uses the Logger instance in its report method, this method could be invoked during the Stats instance destruction, which renders the behavior of the program undefined. Likewise, to support useful behaviors, such as logging previously unreported values during program shutdown, the termination ordering of these singletons must be controlled.
  • When checking for memory leaks: Memory leak detection tools, e.g., NuMega BoundsCheck, ParaSoft Insure++, and Rational Purify, are useful for languages such as C and C++ that require explicit allocation and deallocation of dynamic memory. These tools identify singleton instances as leaked memory, reports of which obscure important memory leaks.
For instance, if many identifiers are used to record Stats data in our running example, it may appear that a sizable amount of memory is leaking during program operation. In large-scale applications, these "leaks" can result in numerous erroneous warnings, thereby obscuring real memory leaks and hindering system debugging.

  • Dynamic memory allocation may be from a global pool: Some real-time OSes, such as VxWorks and pSOS, have only a single, global heap for all applications. Therefore, application tasks must release dynamically allocated memory upon task termination; otherwise, it cannot be reallocated to other applications until the OS is rebooted. Failure to explicitly release memory that was allocated dynamically by the Logger and Stats singletons in our running example represents real resource leaks on such platforms.
Figure 1
Figure 1. The structure and participants of the OLM.

Problems with static object lifetime. Some objects must be created prior to any use. In C++, such instances traditionally have been created as static objects, which are intended to be constructed prior to invocation of the main program entry point and destroyed at program termination. However, there are several important drawbacks to static objects:

  • Unspecified order of construction/destruction. C++ only specifies the order of construction/destruction of static objects within a compilation unit (file); the construction order matches the declaration order, and destruction is the reverse order.5 However, there is no constraint specified on the order of construction/destruction between static objects in different files. Therefore, construction/destruction ordering is implementation-dependent. For example, the versions of the Logger and Stats classes that use the Singleton Destroyer in the subsection Example illustrate problems that arise from the undefined order of destruction of the Stats and Logger singleton instances.
It is hard to write portable C++ code that uses static objects possessing initialization dependencies. Often, it is simpler to avoid using static objects altogether, rather than trying to analyze for, and protect against, such dependencies. This approach is particularly appropriate for reusable components and frameworks6, which should avoid unnecessary constraints on how they are used and/or initialized.

Explicit singleton management is necessary for correct program operation on some platforms because they destroy singletons prematurely. For instance, the garbage collector in older Java Development Kits (JDKs) may destroy an object when there are no longer any references to it, even if the object was intended to be a singleton.7 Though this deficiency has been fixed in later JDKs, the OLM could solve it, under application control, by maintaining singleton references.

Another problem in Java applications is the sharing of namespaces, which allow sharing (intended or otherwise) between singletons in separate applets.8 Again, the OLM can be used to register singleton instances. Applets would then access their singletons from this registry.

  • Poor support by embedded systems. Embedded systems have historically used C. Therefore, they do not always provide seamless support for OO programming language features. For instance, the construction/destruction of static objects in C++ is one such feature that often complicates embedded systems programming. The embedded OS may have support for explicit invocation of static constructor/destructor calls, but this is not optimal from a programmer's perspective.
Some embedded OSes do not support the notion of a program that has a unique entry point. For example, VxWorks supports multiple tasks, which are similar to threads because they all share one address space. However, there is no designated main task for each application. Therefore, these embedded systems platforms can be configured to call static constructors/destructors at module (object file) load/unload time, respectively. On such platforms, it is not otherwise necessary to unload and load between repeated executions. To properly destroy and construct static objects, however, the static object destructors/constructors must either be called manually, or the module unloaded and loaded again, which hinders repeated testing.

In addition, placement of data in read-only memory (ROM) complicates the use of static objects.9 The data must be placed in ROM prior to runtime; however, static constructors are called at runtime. Therefore, embedded systems sometimes do not support calls of static constructors and destructors. Moreover, if they are supported it may be under explicit application control, instead of by implicit arrangement of the compiler and runtime system.

  • Static objects extend application startup time. Static objects may be constructed at application startup time, prior to invocation of the main entry point. If these objects are not used during a specific execution, then application construction (and destruction) times are needlessly extended. One way to eliminate this waste is to replace each such static object with one that is allocated on demand, e.g., by using the Singleton pattern.
Replacement of a static object with a singleton also can be used to delay construction until the first use of the object. Again, this reduces startup time. Some real-time applications may find it advantageous to construct the singletons at a specific time, after the main entry point has been entered, but before the objects are needed.

One or more of these drawbacks of static objects typically provides sufficient motivation for removing them from a program. Often, it is better not to use them in the first place, but to apply the following solution instead.

Solution Define an OLM, which is a singleton that contains a collection of Preallocated Objects and Managed Objects. The OLM is responsible for constructing and destroying the Preallocated Objects at program initialization and termination, respectively. It is further responsible for ensuring all of its Managed Objects are destroyed properly at program termination.

Applicability Use OLM when:

Singletons and other dynamically created objects must be removed without application intervention at program termination. Singleton and other creational patterns do not typically address the question of when the objects they create should be removed, or who should remove them. In contrast, OLM provides a convenient, global object that deletes dynamically created objects. Creational pattern objects can then register with the OLM for deletion, which usually occurs at program termination.

Static objects must be removed from the application. As described in the subsection Problem, static objects can be troublesome, especially in some languages and on some platforms. OLM provides a mechanism to replace static objects with Preallocated Objects. Preallocated Objects are dynamically allocated before the application uses them, and deallocated at program termination.

The platform does not support static object construction/destruction. Some embedded platforms, such as VxWorks and pSOS, do not always construct static objects at program initialization and destroy them at program termination.Ý In general, it is best to remove all static objects, e.g., to support repeated testing of a program. Another situation where static objects can cause difficulty is when they are placed in ROM. Objects in ROM cannot be initialized at runtime, because they cannot be modified at all.

The underlying platform does not provide a notion of a main program, although the application needs it. The lack of support for static object construction/destruction on some platforms stems from their lack of support for the notion of a program, as discussed in the subsection Problem. The OLM pattern can be used to emulate programs by partitioning this address space. The scope of each OLM delineates a program, from an application's perspective.

Destruction order must be specified by the application. Dynamically created objects can be registered with the OLM for destruction. The OLM can be implemented to destroy objects in any desired order.

The application requires explicit singleton management. Singletons may be destroyed prematurely, for example, on earlier Java platforms. The OLM delays singleton destruction until program termination.

Structure and Participants The structure and participants of the OLM pattern are shown using UML in the following figure and described below.

Object Lifetime Manager: Each Object Lifetime Manager is a singleton that contains collections of Managed Objects and Preallocated Objects.

Managed Objects: Any object may be registered with an Object Lifetime Manager, which is responsible for destroying the object. Object destruction occurs when the Object Lifetime Manager itself is destroyed, typically at program termination.

Preallocated Objects: An object may be hard-coded for construction and destruction by an Object Lifetime Manager. Preallocated Objects have the same lifetime as the Object Lifetime Manager, i.e., the lifetime of the process that executes the program.

Application: The Application initializes and destroys its Object Lifetime Managers, either implicitly or explicitly. In addition, the Application registers its Managed Objects with an Object Lifetime Manager, which may also contain Preallocated Objects.

Figure 2
Figure 2. Dynamic collaborations among OLM participants.

Dynamics The dynamic collaborations among participants in the OLM pattern are shown in Figure 2, which depicts four separate activities:

  1. Object Lifetime Manager creation and initialization, which in turn creates the Preallocated Objects;
  2. Managed Object creation by the Application, and registration with the Object Lifetime Manager;
  3. Use of Preallocated and Managed Objects by the Application;
  4. Destruction of the Object Lifetime Manager, which includes destruction of all Managed and Preallocated Objects it controls.
Within each activity, time increases down the vertical axis.

Implementation The OLM pattern can be implemented using the steps presented here. This implementation is based on the Object Manager provided in the ACE framework, which motivates many interesting issues discussed in this section.6 Some of the steps discussed are language-specific because ACE is written in C++.

  • Define the OLM component. This component provides applications with an interface with which to register objects whose lifetime must be managed to ensure proper destruction upon program termination. In addition, this component defines a repository that ensures proper destruction of its managed objects. The Object Lifetime Manager is a container for the Preallocated Objects and for the Managed Objects that are registered to be destroyed at program termination.
The following substeps can be used to implement the Object Lifetime Manager.

  • Define an interface for registering Managed Objects. One way to register Managed Objects with the Object Lifetime Manager would be to use the C-library atexit function to invoke the termination functions at program exit. However, not all platforms support atextit. Furthermore, atexit implementations usually have a limit of 32 registered termination functions. Therefore, the Object Lifetime Manager should support the following two techniques for registering Managed Objects with a container that holds these objects and cleans them up automatically at program exit:

  1. Define a cleanup function interface. The Object Lifetime Manager allows applications to register arbitrary types of objects. When a program is shut down, the Object Lifetime Manager cleans up these objects automatically.

    The following C++ class illustrates a specialized CLEANUP_FUNC used in ACE to register an object or array for cleanup:

    typedef void (*CLEANUP_FUNC)(void *object, void *param);
    
    class Object_Lifetime_Manager
    {
    public:
    	// . . .
    	static int at_exit (void *object,
    			CLEANUP_FUNC cleanup_hook,
    			void *param);
    	// . . .
    };
    The static at_exit method registers an object or array of objects for cleanup at process termination. The cleanup_hook argument points to a global function or static method that is called at cleanup time to destroy the object or array. At destruction time, the Object Lifetime Manager passes the object and param arguments to the cleanup_hook function. The param argument contains any additional information needed by the cleanup_hook function, such as the number of objects in the array.
  2. Define a cleanup base class interface. This interface allows applications to register for destruction with the Object Lifetime Manager any object whose class derives from a Cleanup base class. The Cleanup base class should have a virtual destructor and a cleanup method that simply calls delete this, which in turn invokes all derived class destructors. The following code fragment illustrates how this class is implemented in ACE:

    class Cleanup
    {
    public:
    	// . . .
    
    	// Destructor.
    	virtual ~Cleanup (void);
    
    	// By default, simply deletes this.
    	virtual void cleanup (void *param = 0);
    };
    The following code fragment illustrates the Object Lifetime Manager interface used to register objects derived from the Cleanup base class in ACE:

    class Object_Lifetime_Manager
    {
    public:
    	// . . .
    	static int at_exit (Cleanup *object,
    				void *param = 0);
    	// . . .
    };
    This static at_exit method registers a Cleanup object for cleanup at process termination. At destruction time, the Object Lifetime Manager calls the Cleanup object's cleanup method, passing in the param argument. The param argument contains any additional information needed by the cleanup method.
  • Define a singleton adapter. Although it is possible to explicitly code singletons to use the Object Lifetime Manager methods defined above, this approach is tedious and error-prone. Therefore, it is useful to define a Singleton adapter class template that encapsulates the details of creating singleton objects and registering them with the Object Lifetime Manager. In addition, the Singleton adapter can ensure the thread-safe Double-Checked Locking Optimization pattern is used to construct and access an instance of the type-specific Singleton.10
The following code fragment illustrates how a singleton adapter is implemented in ACE:

template 
class Singleton : public Cleanup
{
public:
	// Global access point to the
	  // wrapped singleton.
	static TYPE *instance (void) {
		// Details of Double Checked
		// Locking Optimization omitted . . .
		if (singleton_ == 0) {
			singleton_ = new Singleton;

			// Register with the Object Lifetime
			// Manager to control destruction.
			Object_Lifetime_Manager::
				at_exit (singleton_);
		}
		return &singleton_->instance_;
	}

	protected:
	// Default constructor.
	Singleton (void);

	// Contained instance.
	TYPE instance_;

	// Instance of the singleton adapter.
	static Singleton *singleton_;
};
The Singleton class template is derived from the Cleanup class. This allows the Singleton instance to register itself with the Object Lifetime Manager. The Object Lifetime Manager then assumes responsibility for dynamically deallocating the Singleton instance and with it the adapted TYPE instance.

  • Define an interface for registering Preallocated Objects. Preallocated Objects must always be created before the application's main processing begins. For instance, in some applications, synchronization locks must be created before their first use to avoid race conditions. Thus, these objects must be hard-coded into each Object Lifetime Manager class. Encapsulating their creation within the Object Lifetime Manager's own initialization phase ensures this will occur, without adding complexity to application code.
The Object Lifetime Manager should be able to preallocate objects or arrays. The Object Lifetime Manager can either perform these preallocations statically in global data or dynamically on the heap. An efficient implementation is to store each Preallocated Object in an array. Certain languages, such as C++, do not support arrays of heterogeneous objects, however. Therefore, in these languages pointers must be stored instead of the objects themselves. The actual objects are allocated dynamically by the Object Lifetime Manager when it is instantiated, and destroyed by the Object Lifetime Manager destructor.

The following substeps should be used to implement Preallocated Objects:

  1. Limit exposure. To minimize the exposure of header files, identify the Preallocated Objects by macros or enumerated literals, for example:
    enum Preallocated_Object_ID
    	{
    		ACE_FILECACHE_LOCK,
    		ACE_STATIC_OBJECT_LOCK,
    		ACE_LOG_MSG_INSTANCE_LOCK,
    		ACE_DUMP_LOCK,
    		ACE_SIG_HANDLER_LOCK,
    		ACE_SINGLETON_NULL_LOCK,
    		ACE_SINGLETON_RECURSIVE_THREAD_LOCK,
    		ACE_THREAD_EXIT_LOCK,
    	};
    The use of an enumerated type is appropriate when all preallocated objects are known a priori.
  2. Use cleanup adapters. The Cleanup Adapter class template is derived from the Cleanup base class and wraps types not derived from the Cleanup base class so that they can be managed by the Object Lifetime Manager. Use template functions or macros for allocation/deallocation, for example:

    #define PREALLOCATE_OBJECT(TYPE, ID) {	Cleanup_Adapter *obj_p;	obj_p = new Cleanup_Adapter;	preallocated_object[ID] = obj_p;}
    
    	#define DELETE_PREALLOCATED_OBJECT(TYPE, ID)	cleanup_destroyer (		static_cast<> *>			(preallocated_object[ID]), 0);	preallocated_object[ID] = 0;
    The Cleanup Adapter adapts any object to use the simpler Object Lifetime Manager registration interface. Similarly, Cleanup Destroyer uses the Cleanup Adapter to destroy the object.

    An analogous array, enum, and macro pair can be supplied for preallocated arrays, if necessary.

  3. Define an accessor interface to the Preallocated Objects. Applications need a convenient and type-safe interface to access preallocated objects. Because the Object Lifetime Manager supports preallocated objects of different types, it is necessary to provide a separate class template adapter with a member function that takes the id under which the object was preallocated in the Object Lifetime Manager, and returns a correctly typed pointer to the preallocated object.

    The following code fragment illustrates how this interface is provided via a class template adapter in ACE:

    template 
    class Preallocated_Object_Interface
    {
    public:
    	static TYPE *
    	get_preallocated_object
    		(Object_Lifetime_Manager::
    			Preallocated_Object_ID id)
    	{
    		// Cast the return type of the object
    		// pointer based on the type of the
    		// function template parameter.
    		return
    			&(static_cast<> *>
    				(Object_Lifetime_Manager::
    					preallocated_object[id])->object ());
    	}
    	// . . . other methods omitted.
    };
  • Determine the destruction order of registered objects. As noted in the subsection Applicability, the Object Lifetime Manager can be implemented to destroy registered objects in any desired order. For example, priority levels could be assigned and destruction could proceed in decreasing order of priority. An interface could be provided for objects to set and change their destruction priority.
We have found that destruction in reverse order of registration has been a sufficient policy for ACE applications. An application can, in effect, specify destruction order by controlling the order in which it registers objects with the Object Lifetime Manager.

  • Define a termination function interface. Lifetime management functionality has been discussed so far in terms of destruction of objects at program termination. However, the Object Lifetime Manager can provide a more general capability-the ability to call a function at program termination-using the same internal implementation mechanism.
For example, to ensure proper cleanup of open Win32 WinSock sockets at program termination, the WSACleanup function must be called. A variation of the Scoped Locking C++ idiom4 could be applied, by creating a special wrapper facade11 class whose constructor calls the API initialization functions, and whose destructor calls the API cleanup functions. The application can then register an instance of this class with the Object Lifetime Manager so that the object is destroyed and API termination methods are called during Object Lifetime Manager termination.

However, this design may be overkill for many applications. Moreover, it can be error-prone because the application must ensure the class is used only as a singleton so that the API functions are only called once. Finally, this design may increase the burden on the application to manage object destruction order, so that the singleton is not destroyed before another object that uses the API in its destructor.

Instead, the API termination functions can be called as part of the termination method of the Object Lifetime Manager itself. This latter approach is illustrated by the socket_fini call in the Object_Lifetime_Manager::fini method. The socket_fini function provides a platform-specific implementation that calls WSACleanup on Win32 platforms, and does nothing on other platforms.

  • Migrate common interfaces and implementation details into a base class. Factoring common internal details into an Object Lifetime Manager Base class can make the Object Lifetime Manager implementation simpler and more robust. Defining an Object Lifetime Manager Base class also supports the creation of multiple Object Lifetime Managers, each of a separate type. To simplify our discussion, we only touch on the use of multiple Object Lifetime Managers briefly. They do not add consequences to the pattern, but are useful for partitioning libraries and applications.
  • Determine how to manage the lifetime of the OLM itself. The Object Lifetime Manager is responsible for initializing other global and static objects in a program. However, that begs the important bootstrapping question of how this singleton initializes and destroys itself. The following are the alternatives for initializing the Object Lifetime Manager singleton instance:
  • Static initialization. If an application has no static objects with constraints on their order of construction or destruction, it's possible to create the Object Lifetime Manager as a static object. For example, ACE's Object Lifetime Manager can be created as a static object. The ACE library has no other static objects that have constraints on order of construction or destruction.
When there is a main program thread with well-defined points of program entry and termination, creating the Object Lifetime Manager on the stack of the main program thread can simplify program logic for creating and destroying the Object Lifetime Manager. This approach to initializing the Object Lifetime Manager assumes that there is one unique main thread per program. This thread defines the program: i.e., it is running if, and only if, the main thread is alive. This approach has a compelling advantage: The Object Lifetime Manager instance is automatically destroyed via any path out of main.ý

Stack initialization is implemented transparently in ACE via a preprocessor macro named main. The macro renames the main program entry point to another, configurable name, such as main_i. It provides a main function that creates the Object Lifetime Manager instance as a local object (on the runtime stack) and then calls the renamed application entry point.

There are two drawbacks to the Stack initialization approach:

  1. main (int, char *[]) must be declared with arguments, even if they're not used. All of ACE uses this convention, so only applications must be concerned with it.
  2. If there are any static objects depending on those that are destroyed by the Object Lifetime Manager, their destructors might attempt to access the destroyed objects. Therefore, the application developer is responsible for ensuring that no static objects depend on those destroyed by the Object Lifetime Manager.
  • Explicit initialization. In this approach, create the Object Lifetime Manager explicitly under application control. The Object Lifetime Manager init and fini methods allow the application to create and destroy the OLM when desired. This option alleviates complications that arise when using dynamic link libraries (DLL)s.
  • Dynamic link library initialization. In this approach, create and destroy the OLM when its DLL is loaded and unloaded, respectively. Most dynamic library facilities include the ability to call 1) an initialization function when the library is loaded and 2) a termination function when the library is unloaded. We give a detailed discussion of the consequences of managing singletons from an OLM in different DLLs later in this article.
Example Resolved The discussion in the previous subsection demonstrates how the OLM pattern can be used to resolve key design forces related to managing object lifetimes. In particular, the unresolved forces described in the Example subsection can be satisfied by applying the OLM pattern.

The following example shows how the OLM pattern can be applied to the original Logger and Stats examples. Using the Singleton adapter template described here greatly simplifies managed object implementations by encapsulating key implementation details, such as registration with the OLM. For instance, the original Stats class can be replaced by a managed Stats class, as follows:

class Stats
{
public:
	friend class Singleton;

	// Destructor: frees resources.
	~Stats (void);

	// Record a timing data point.
	int record (int id, const timeval &tv);

	// Report recorded statistics
	// to the log.
	void report (int id) {
		Singleton::instance ()->
			log ("Avg timing %d: "
				"%ld sec %ld usec\n",
				id,
				average_i (id).tv_sec,
				average_i (id).tv_usec);
	}

protected:
	// Default constructor.
	Stats (void)
	{
		// Ensure the Logger instance
		// is registered first, and will be
		// cleaned up after, the Stats
		// instance.
		Singleton::instance ();
	}

	// Internal accessor for an average.
	const timeval &average_i (void);

	// . . . other resources that are
	//       held by the instance . . .
};
Notice that the singleton aspects have been factored out of the original Stats class and are now provided by the Singleton adapter template. Similar modifications can be made to the original Logger class so that it uses the Singleton adapter template.

Finally, the following example shows how an application might use the Logger and Stats classes.

int main (int argc, char *argv[])
{
	// Interval timestamps.
	timeval start_tv, stop_tv;

	// Logger and Stats singletons
	// do not yet exist.

	// Logger and then Stats singletons
	// are created and registered on the first
	// iteration.
	for (int i = 0; i < argc;="" ++i)="" {="" ::gettimeofday="" (&start_tv);="" do="" some="" work="" between="" timestamps="" ...="" ::gettimeofday="" (&stop_tv);="" then="" record="" the="" stats="" ...="" timeval="" delta_tv;="" delta_tv.sec="stop_tv.sec" -="" start_tv.sec;="" delta_tv.usec="stop_tv.usec" -="" start_tv.usec;="">::instance ()->
			record (i, delta_tv);

		// . . . and log some output.
		Singleton::instance ()->
			log ("Arg %d [%s]\n", i, argv[i]);
		Singleton::instance()->report (i);
	}

	// Logger and Stats singletons are
	// cleaned up by OLM
	// upon program exit.
	return 0;
}
The following key forces are resolved in this example: 1) ensuring resources allocated by an instance are subsequently released, 2) managing the order of creation and destruction of singletons, and 3) providing a framework that encapsulates these details within a well-defined interface.

Known Uses OLM is used in the Adaptive Communication Environment (ACE)6 to ensure destruction of singletons at program termination and to replace static objects with dynamically allocated, managed objects. ACE is used on many different OS platforms, some of which do not support static object construction/destruction for every program invocation. ACE can be configured to not contain any objects whose initialization is necessary prior to program invocation.

Gabrilovich12 augmented the Singleton pattern to permit applications to specify destruction order. A local static auto_ptr is responsible for the destruction of each singleton instance.# Destruction of singleton instances proceeds by application-defined phases; an application may optionally register its singleton instances for destruction in a specific phase.

An interesting example of a "small" OLM is the strong pointer13, 14.|| A strong pointer manages just one object; it destroys the object when its scope is exited, either normally or via an exception. There can be many strong pointers in a program, behaving as Function-as-Owner (or Block-as-Owner). Moreover, the strong pointers themselves have transient lifetimes, i.e., that of their enclosing blocks.

In contrast, there is typically just one OLM per program (or per large-scale component). And, OLMs live for the duration of the program invocation. This reflects the specific intent of the OLM to destroy objects at program termination, but not sooner. Such objects may be used after the current block or function has been exited, and destruction/creation cycles are not possible or desired.

See Also The OLM pattern is related to the Manager pattern.15 In both patterns, a client application uses a collection of objects, relying upon a manager to encapsulate the details of how the objects themselves are managed. The OLM pattern differs from the Manager pattern in the types of the managed objects. Whereas the Manager pattern requires that the managed objects have a common base type, the OLM pattern allows objects of unrelated types to be managed. The Manager pattern relies on inheritance for variations in the manager and managed object classes. In contrast, the OLM relies on object composition and type parameterization to achieve greater decoupling of the manager from the managed objects.

OLM complements creational patterns such as Singleton, by managing object instance destruction. Singleton addresses only part of the object lifetime because it just manages instance creation. However, destruction is usually not an important issue with Singleton because it does not retain ownership of created objects.3 Ownership conveys the responsibility for managing the object, including its destruction. Singleton is the prototypical example of a creational pattern that does not explicitly transfer ownership, yet does not explicitly arrange for object destruction. OLM complements Singleton by managing the destruction portion of the object lifetime.

OLM can complement other creational patterns, such as Abstract Factory and Factory Method. Implementations of these patterns could register dynamically allocated objects for deletion at program termination. Alternatively (or additionally), they could provide interfaces for object destruction, corresponding to those for object creation.

Cargill presented a taxonomy of the dynamic C++ object lifetime.16 The Localized Ownership pattern language includes patterns, such as Creator-as-Owner, Sequence-of-Owners, and Shared Ownership, which primarily address object ownership. Ownership conveys the responsibility for destruction.

Creator-as-Owner is further subdivided into Function-as-Owner, Object-as-Owner, and Class-as-Owner. The Singleton destruction capability of OLM may be viewed as new category of Creator-as-Owner: Program-as-Owner. It is distinct from Function-as-Owner, because static objects outlive the program entry point (main). OLM's Preallocated Objects similarly can be viewed logically, at least, as outliving the main program function.

CONSEQUENCES The benefits of using OLM include:

Destruction of Singletons and other Managed Objects at program termination. The OLM pattern allows a program to shut down cleanly, releasing memory for Managed Objects, along with the resources they hold at program termination. All heap-allocated memory can be released by the application. This supports repeated testing on platforms where heap allocations outlive the program.** It also eliminates the memory-in-use warnings reported for singletons by memory access checkers at program termination.

Specification of destruction order. The order of destruction of objects can be specified. The order specification mechanism can be as simple or as complex as desired. As noted in the subsection Implementation, simple mechanisms are generally sufficient in practice.

Removal of static objects from libraries and applications. Static objects can be replaced by Preallocated Objects. This prevents applications from relying on the order in which static objects are constructed/destructed. Moreover, it allows code to target embedded systems, which sometimes have little or no support for constructing/destroying static objects.

However, the following liabilities must be considered when using the OLM pattern:

Lifetime of the manager itself. The application must ensure that it respects the lifetime of the OLM, and does not attempt to use its services outside that lifetime. For example, the application must not attempt to access Preallocated Objects prior to the complete initialization of the Object Lifetime Manager. Similarly, the application must not destroy the Object Lifetime Manager prior to the application's last use of a Managed or Preallocated Object. Finally, the implementation of the Object Lifetime Manager is simplified if it can assume that it will be initialized by only one thread. This precludes the need for a static lock to guard its initialization.

Use with shared libraries. On platforms that support loading and unloading shared libraries at runtime, the application must be very careful of platform-specific issues that impact the lifetime of the Object Lifetime Manager itself. For example, on Windows NT, the Object Lifetime Manager should be initialized by the application or by a DLL that contains it. This avoids a potential deadlock situation due to serialization within the OS when it loads DLLs.

A related issue arises with singletons that are created in DLLs but managed by an Object Lifetime Manager in the main application code. If the DLL is unloaded before program termination, the Object Lifetime Manager will try to destroy it using code that is no longer linked into the application. For this reason, we have added an unmanaged Singleton class to ACE. An unmanaged Singleton is of the conventional design, i.e., it does not provide implicit destruction. ACE uses a managed Singleton by default because we found the need for unmanaged Singletons to be very unusual.

CONCLUSION Many creational patterns specifically address only object creation. They do not consider when or how to destroy objects that are no longer needed. The OLM pattern provides mechanisms for object destruction at program termination. Thus, it complements many creational patterns by covering the entire object lifetime.

The Singleton pattern provides a notable example where coordinated object lifetime management is important. In particular, deletion at program termination ensures that programs have no memory leaks of singleton objects. Moreover, applications that employ the OLM pattern do not require use of static object constructors and destructors, which is important for embedded systems. In addition, the OLM pattern supports replacement of static objects with dynamically Preallocated Objects, which is useful on embedded platforms and with OO languages such as C++.

One of the OLM pattern's more interesting aspects is that it addresses weaknesses of another pattern, at least in some contexts. Our initial motivation was to remedy these weaknesses by registering and deleting singletons at program termination. The utility, applicability, novelty, and complexity of the ACE Object Lifetime Manager class seemed to be on par with those of the ACE Singleton adapter template class, so we felt that it deserved consideration as a pattern. Because it can address just part of the object lifetime, however, we consider OLM to be a complementary pattern.

A highly portable implementation of the OLM pattern and the Singleton adapter template is freely available and can be downloaded.17 We hope that you have gained a greater appreciation for the motivations for avoiding static objects in portable C++ programs, and have added new techniques to your toolbox for programming without them.

Acknowledgments This work was supported in part by Boeing, NSF grant NCR-9628218, DARPA contract 9701516, and Nortel.

Thanks to Matthias Kerkhoff, Per Andersson, Steve Huston, Elias Sreih, and Liang Chen for many helpful discussions on the design and implementation of ACE's Object Manager. Thanks to Brad Appleton, our PLoP '99 shepherd and C++ Report Patterns++ section editor, Evgeniy Gabrilovich, Kevlin Henney, and our PLoP '99 workshop group for many helpful suggestions on the content and presentation of the OLM pattern. And thanks to Bosko Zivaljevic for pointing out that static object construction can extend application startup time.

References

  1. Gamma, E., et.al. Design Patterns: Elements of Reusable Object-Oriented Software, Addison-Wesley, Reading, MA, 1995.
  2. Schmidt, D. C. "Reactor: An Object Behavioral Pattern for Concurrent Event Demultiplexing and Event Handler Dispatching," in Pattern Languages of Program Design, J.O. Coplien and D.C. Schmidt, Eds., pp. 529-545, Addison-Wesley, Reading, MA, 1995.
  3. Vlissides, J. Pattern Hatching: Design Patterns Applied, Addison-Wesley, Reading, MA, 1998.
  4. Schmidt, D. C. "Strategized Locking, Thread-safe Decorator, and Scoped Locking: Patterns and Idioms for Simplifying Multi-threaded C++ Components," C++ Report 11(8), Sept. 1999.
  5. Stroustrup, B. The C++ Programming Language, 3rdEdition, Addison-Wesley, Reading, MA, 1998.
  6. Schmidt, D. C. "ACE: an Object-Oriented Framework for Developing Distributed Applications," in Proceedings of the 6thUSENIX C++ Technical Conference, (Cambridge, Massachusetts), USENIX Association, April 1994.
  7. "Java Singleton." http://c2.com/cgi/wiki?JavaSingleton, May 1999.
  8. "Another Java Singleton Problem." http://c2.com/cgi/wiki? Another Java Singleton Problem, Jan. 1999.
  9. Saks, D. "Ensuring Static Initialization in C++," Embedded Systems Programming, vol.12, pp. 109-111, Mar. 1999.
  10. Schmidt, D. C., and T. Harrison, "Double-Checked Locking - An Object Behavioral Pattern for Initializing and Accessing Thread-safe Objects Efficiently," in Pattern Languages of Program Design (R. Martin, F. Buschmann, and D. Riehle, Eds.), Addison-Wesley, Reading, MA, 1997.
  11. Schmidt, D. C. "Wrapper Facade: A Structural Pattern for Encapsulating Functions within Classes," C++ Report, 11(2), Feb. 1999.
  12. Gabrilovich, E. "Destruction-Managed Singleton: A Compound Pattern for Reliable Deallocation of Singletons," C++ Report, 12(1), Feb. 2000.
  13. Milewski, B. "Strong Pointers and Resource Management in C++," C++ Report, 10(8): 23-27, Sept. 1998.
  14. Milewski, B. "Strong Pointers and Resource Management in C++, Part 2," C++ Report, 11(2): 36-39, 50, Feb. 1999.
  15. Sommerland, P., and F. Buschmann, "The Manager Design Pattern," in Proceedings of the 3rd Pattern Languages of Programming Conference, September 1996.
  16. Cargill, T. "Localized Ownership: Managing Dynamic Objects in C++," in Pattern Languages of Program Design 2 (J. Vlissides, J. O. Coplien, and N. L. Kerth, Eds.), Addison-Wesley, Reading, MA, 1996.
  17. Object Life Manager Pattern and Singleton adapter template. www.cs.wustl.edu/~schmidt/ACE-obtain.html.
FOOTNOTES
* Singleton is used as an example in much of this pattern description because 1) it is a popular creational pattern and 2) it highlights challenging object destruction issues nicely. However, OLM can complement other creational patterns, such as Factory Method, and does not assume that its managed objects are singletons or of homogeneous types.

Ý On VxWorks and pSOS, static objects can be constructed when the module is loaded and destroyed when it is unloaded. After loading, an entry point can be called more than once before unloading. Therefore, a program can be run more than once after constructing static objects, without ever destroying them. Conversely, static object constructors and destructors can be invoked explicitly.

ý On platforms such as VxWorks and pSOS that have no designated main function, the main thread can be simulated by instantiating the Object Lifetime Manager on the stack of one thread, which is denoted by convention as the main thread.

# The auto_ptr is a local static object in the singleton instance accessor method.

|| A C++ auto_ptr is an implementation of a strong pointer.

** On some operating systems, notably some real-time operating systems, there is no concept of a program. There are tasks, i.e., threads, but no one task has any special, main identity. Thus, there is no cleanup of dynamically allocated memory, open file, etc., at task termination.