what is the correct way to implement a QThread... (example please...)

Not to detract from @sergey-tachenov's excellent answer, but in Qt5 you can stop using SIGNAL and SLOT, simplify your code and have the advantage of compile time checking:

void ChildProcesses::start()
{
    QThread *childrenWatcherThread = new QThread();
    ChildrenWatcher *childrenWatcher = new ChildrenWatcher();
    childrenWatcher->moveToThread(childrenWatcherThread);
    // These three signals carry the "outcome" of the worker job.
    connect(childrenWatcher, ChildrenWatcher::exited,
            ChildProcesses::onChildExited);
    connect(childrenWatcher, ChildrenWatcher::signalled,
            ChildProcesses::onChildSignalled);
    connect(childrenWatcher, ChildrenWatcher::stateChanged,
            ChildProcesses::onChildStateChanged);
    // Make the watcher watch when the thread starts:
    connect(childrenWatcherThread, QThread::started,
            childrenWatcher, ChildrenWatcher::watch);
    // Make the watcher set its 'stop' flag when we're done.
    // This is performed while the watch() method is still running,
    // so we need to execute it concurrently from this thread,
    // hence the Qt::DirectConnection. The stop() method is thread-safe
    // (uses a mutex to set the flag).
    connect(this, ChildProcesses::stopped,
            childrenWatcher, ChildrenWatcher::stop, Qt::DirectConnection);
    // Make the thread quit when the watcher self-destructs:
    connect(childrenWatcher, ChildrenWatcher::destroyed,
            childrenWatcherThread, QThread::quit);
    // Make the thread self-destruct when it finishes,
    // or rather, make the main thread delete it:
    connect(childrenWatcherThread, QThread::finished,
            childrenWatcherThread, QThread::deleteLater);
    childrenWatcherThread->start();
}

About the only thing I can think of to add is to further state that QObjects have an affinity with a single thread. This is usually the thread that creates the QObject. So if you create a QObject in the app's main thread and want to use it in another thread, you need to use moveToThread() to change the affinity.

This saves having to subclass QThread and creating your objects in the run() method, thus keeping your stuff nicely encapsulated.

That blog post does include a link to an example. It is pretty short but it shows the basic idea. Create your QObjects, connect your signals, create your QThread, move your QObjects to the QThread and start the thread. The signal/slot mechanisms will ensure that thread boundaries are crossed properly and safely.

You may have to introduce synchronization if you have to call methods on your object outside of that mechanism.

I know Qt has some other nice threading facilities beyond threads that are probably worth getting familiar with but I have yet to do so :)

Here's one example of how to use QThread correctly, but it has some issues with it, which are reflected in the comments. In particular, since the order in which the slots are executed isn't strictly defined, it could lead to various problems. The comment posted on August 6, 2013 gives a nice idea how to deal with this issue. I use something like that in my program, and here's some example code to clarify.

The basic idea is the same: I create a QThread instance that lives in my main thread, a worker class instance that lives in the new thread I created, and then I connect all the signals.

void ChildProcesses::start()
{
    QThread *childrenWatcherThread = new QThread();
    ChildrenWatcher *childrenWatcher = new ChildrenWatcher();
    childrenWatcher->moveToThread(childrenWatcherThread);
    // These three signals carry the "outcome" of the worker job.
    connect(childrenWatcher, SIGNAL(exited(int, int)),
            SLOT(onChildExited(int, int)));
    connect(childrenWatcher, SIGNAL(signalled(int, int)),
            SLOT(onChildSignalled(int, int)));
    connect(childrenWatcher, SIGNAL(stateChanged(int)),
            SLOT(onChildStateChanged(int)));
    // Make the watcher watch when the thread starts:
    connect(childrenWatcherThread, SIGNAL(started()),
            childrenWatcher, SLOT(watch()));
    // Make the watcher set its 'stop' flag when we're done.
    // This is performed while the watch() method is still running,
    // so we need to execute it concurrently from this thread,
    // hence the Qt::DirectConnection. The stop() method is thread-safe
    // (uses a mutex to set the flag).
    connect(this, SIGNAL(stopped()),
            childrenWatcher, SLOT(stop()), Qt::DirectConnection);
    // Make the thread quit when the watcher self-destructs:
    connect(childrenWatcher, SIGNAL(destroyed()),
            childrenWatcherThread, SLOT(quit()));
    // Make the thread self-destruct when it finishes,
    // or rather, make the main thread delete it:
    connect(childrenWatcherThread, SIGNAL(finished()),
            childrenWatcherThread, SLOT(deleteLater()));
    childrenWatcherThread->start();
}

Some background:

The ChildProcesses class is a child process manager that starts new child processes with spawn() calls, keeps the list of the processes currently running and so on. However, it needs to keep track of the children states, which means using waitpid() call on Linux or WaitForMultipleObjects on Windows. I used to call these in non-blocking mode using a timer, but now I want more prompt reaction, which means blocking mode. That's where the thread comes in.

The ChildrenWatcher class is defined as follows:

class ChildrenWatcher: public QObject {
    Q_OBJECT
private:
    QMutex mutex;
    bool stopped;
    bool isStopped();
public:
    ChildrenWatcher();
public slots:
    /// This is the method which runs in the thread.
    void watch();
    /// Sets the stop flag.
    void stop();
signals:
    /// A child process exited normally.
    void exited(int ospid, int code);
    /// A child process crashed (Unix only).
    void signalled(int ospid, int signal);
    /// Something happened to a child (Unix only).
    void stateChanged(int ospid);
};

Here how it works. When all this stuff is started, the ChildProcess::start() method is called (see above). It creates a new QThread and a new ChildrenWatcher, which is then moved to the new thread. Then I connect three signals which inform my manager about the fate of its child processes (exited/signalled/god-knows-what-happened). Then starts the main fun.

I connect QThread::started() to the ChildrenWatcher::watch() method so it is started as soon as the thread is ready. Since the watcher lives in the new thread, that's where the watch() method is executed (queued connection is used to call the slot).

Then I connect the ChildProcesses::stopped() signal to the ChildrenWatcher::stop() slot using Qt::DirectConnection because I need to do it asynchronously. This is needed so my thread stops when the ChildProcesses manager is no longer needed. The stop() method looks like this:

void ChildrenWatcher::stop()
{
    mutex.lock();
    stopped = true;
    mutex.unlock();
}

And then ChildrenWatcher::watch():

void ChildrenWatcher::watch()
{
  while (!isStopped()) {
    // Blocking waitpid() call here.
    // Maybe emit one of the three informational signals here too.
  }
  // Self-destruct now!
  deleteLater();
}

Oh, and the isStopped() method is just a convenient way to use a mutex in the while() condition:

bool ChildrenWatcher::isStopped()
{
    bool stopped;
    mutex.lock();
    stopped = this->stopped;
    mutex.unlock();
    return stopped;
}

So what happens here is that I set the stopped flag when I need to finish, and then the next time isStopped() is called it returns false and the thread ends.

So what happens when the watch() loop ends? It calls deleteLater() so the object self-destructs as soon as control is returned to the thread event loop which happens just right after the deleteLater() call (when watch() returns). Going back to ChildProcesses::start(), you can see that there is a connection from the destroyed() signal of the watcher to the quit() slot of the thread. This means that the thread automatically finishes when the watcher is done. And when it's finished, it self-destructs too because its own finished() signal is connected to its deleteLater() slot.

This is pretty much the same idea as Maya posted, but because I use the self-destruct idiom, I don't need to depend on the sequence in which the slots are called. It's always self-destruct first, stop thread later, then it self-destructs too. I could define a finished() signal in the worker, and then connect it to its own deleteLater(), but that would only mean one connection more. Since I don't need a finished() signal for any other purpose, I chose to just call deleteLater() from the worker itself.

Maya also mentions that you shouldn't allocate new QObjects in the worker's constructor because they won't live in the thread you move the worker to. I'd say do it anyway because that's the way OOP works. Just make sure all those QObjects are children of the worker (that is, use the QObject(QObject*) constructor) - moveToThread() moves all the children along with the object being moved. If you really need to have QObjects that aren't children of your object, then override moveToThread() in your worker so it moves all the necessary stuff too.