The server-side Ice run time by default creates a thread pool for you and automatically dispatches each incoming request in its own thread. As a result, you usually only need to worry about synchronization among threads to protect critical regions when you implement a server. However, you may wish to create threads of your own. For example, you might need a dedicated thread that responds to input from a user interface. And, if you have complex and long-running operations that can exploit parallelism, you might wish to use multiple threads for the implementation of that operation.

Ice provides a simple thread abstraction that permits you to write portable source code regardless of the native threading platform. This shields you from the native underlying thread APIs and guarantees uniform semantics regardless of your deployment platform.

On this page:

The C++ Thread Class

The basic thread abstraction in Ice is provided by two classes, ThreadControl and Thread (defined in IceUtil/Thread.h):

{zcode:cpp}
namespace IceUtil {

    class Time;

    class ThreadControl {
    public:
#ifdef _WIN32
        typedef DWORD ID;
#else
        typedef pthread_t ID;
#endif

        ThreadControl();
#ifdef _WIN32
        ThreadControl(HANDLE, DWORD);
#else
        ThreadControl(explicit pthread_t);
#endif
        ID id() const;

        void join();
        void detach();

        static void sleep(const Time&);
        static void yield();

        bool operator==(const ThreadControl&) const;
        bool operator!=(const ThreadControl&) const;

    };

    class Thread :virtual public Shared {
    public:
        virtual void run() = 0;

        ThreadControl start(size_t stBytes = 0);
        ThreadControl start(size_t stBytes, int priority);
        ThreadControl getThreadControl() const;
        bool isAlive() const;

        bool operator==(const Thread&) const;
        bool operator!=(const Thread&) const;
        bool operator<(const Thread&) const;
    };
    typedef Handle<Thread> ThreadPtr;
}
{zcode}

The Thread class is an abstract base class with a pure virtual run method. To create a thread, you must specialize the Thread class and implement the run method (which becomes the starting stack frame for the new thread). Note that you must not allow any exceptions to escape from run. The Ice run time installs an exception handler that calls ::std::terminate if run terminates with an exception.

The remaining member functions behave as follows:

Note that IceUtil also defines the type ThreadPtr. This is the usual reference-counted smart pointer to guarantee automatic clean-up: the Thread destructor calls delete this once its reference count drops to zero.

Implementing Threads in C++

To illustrate how to implement threads, consider the following code fragment:

{zcode:cpp}
#include <IceUtil/Thread.h>
// ...

Queue q;

class ReaderThread : public IceUtil::Thread {
    virtual void run() {
        for (int i = 0; i < 100; ++i)
            cout << q.get() << endl;
    }
};

class WriterThread : public IceUtil::Thread {
    virtual void run() {
        for (int i = 0; i < 100; ++i)
            q.put(i);
    }
};
{zcode}

This code fragment defines two classes, ReaderThread and WriterThread, that inherit from IceUtil::Thread. Each class implements the pure virtual run method it inherits from its base class. For this simple example, a writer thread places the numbers from 1 to 100 into an instance of the thread-safe Queue class we defined in our discussion of monitors, and a reader thread retrieves 100 numbers from the queue and prints them to stdout.

Creating Threads in C++

To create a new thread, we simply instantiate the thread and call its start method:

{zcode:cpp}
IceUtil::ThreadPtr t = new ReaderThread;
t->start();
// ...
{zcode}

Note that we assign the return value from new to a smart pointer of type ThreadPtr. This ensures that we do not suffer a memory leak:

  1. When the thread is created, its reference count is set to zero.
  2. Prior to calling run (which is called by the start method), start increments the reference count of the thread to 1.
  3. For each ThreadPtr for the thread, the reference count of the thread is incremented by 1, and for each ThreadPtr that is destroyed, the reference count is decremented by 1.
  4. When run completes, start decrements the reference count again and then checks its value: if the value is zero at this point, the Thread object deallocates itself by calling delete this; if the value is non-zero at this point, there are other smart pointers that reference this Thread object and deletion happens when the last smart pointer goes out of scope.

Note that, for all this to work, you must allocate your Thread objects on the heap — stack-allocated Thread objects will result in deallocation errors:

{zcode:cpp}
ReaderThread thread;
IceUtil::ThreadPtr t = &thread; // Bad news!!!
{zcode}

This is wrong because the destructor of t will eventually call delete, which has undefined behavior for a stack-allocated object.

Similarly, you must use a ThreadPtr for an allocated thread. Do not attempt to explicitly delete a thread:

{zcode:cpp}
Thread* t = new ReaderThread();

// ...

delete t; // Disaster!
{zcode}

This will result in a double deallocation of the thread because the thread's destructor will call delete this.

It is legal for a thread to call start on itself from within its own constructor. However, if so, the thread must not be (very) short lived:

{zcode:cpp}
class ActiveObject : public Thread() {
public:
    ActiveObject() {
        start();
    }

    void done() {
        getThreadControl().join();
    }

    virtual void run() {
        // *Very* short lived...
    }
};
typedef Handle<ActiveObject> ActiveObjectPtr;

// ...

ActiveObjectPtr ao = new ActiveObject;
{zcode}

With this code, it is possible for run to complete before the assignment to the smart pointer ao completes; in that case, start will call delete this; before it returns and ao ends up deleting an already-deleted object. However, note that this problem can arise only if run is indeed very short-lived and moreover, the scheduler allows the newly-created thread to run to completion before the assignment of the return value of operator new to ao takes place. This is highly unlikely to happen — if you are concerned about this scenario, do not call start from within a thread's own constructor. That way, the smart pointer is assigned first, and the thread started second, so the problem cannot arise.

The C++ ThreadControl Class

The start method returns an object of type ThreadControl. The member functions of ThreadControl behave as follows:

As for all the synchronization primitives, you must adhere to a few rules when using threads to avoid undefined behavior:

C++ Thread Example

Following is a small example that uses the Queue class we defined in our discussion of monitors. We create five writer and five reader threads. The writer threads each deposit 100 numbers into the queue, and the reader threads each retrieve 100 numbers and print them to stdout:

{zcode:cpp}
#include <vector>
#include <IceUtil/Thread.h>
// ...

Queue q;

class ReaderThread : public IceUtil::Thread {
    virtual void run() {
        for (int i = 0; i < 100; ++i)
            cout << q.get() << endl;
    }
};

class WriterThread : public IceUtil::Thread {
    virtual void run() {
        for (int i = 0; i < 100; ++i)
            q.put(i);
    }
};

int
main()
{
    vector<IceUtil::ThreadControl> threads;
    int i;

    // Create five reader threads and start them
    //
    for (i = 0; i < 5; ++i) {
        IceUtil::ThreadPtr t = new ReaderThread;
        threads.push_back(t->start());
    }

    // Create five writer threads and start them
    //
    for (i = 0; i < 5; ++i) {
        IceUtil::ThreadPtr t = new WriterThread;
        threads.push_back(t->start());
    }

    // Wait for all threads to finish
    //
    for (vector<IceUtil::ThreadControl>::iterator i = threads.begin();
         i != threads.end(); ++i) {
        i->join();
    }
}
{zcode}

The code uses the threads variable, of type vector<IceUtil::ThreadControl>, to keep track of the created threads. The code creates five reader and five writer threads, storing the ThreadControl object for each thread in the threads vector. Once all the threads are created and running, the code joins with each thread before returning from main.

Note that you must not leave main without first joining with the threads you have created: many threading libraries crash if you return from main with other threads still running. (This is also the reason why you must not terminate a program without first calling Communicator::destroy; the destroy implementation joins with all outstanding threads before it returns.)

See Also