How to implement the observer pattern safely? - c++

Question

I'm implementing a mechanism similar to the observer design pattern for a multithreaded Tetris game. There is a Game class which contains a collection of EventHandler objects. If a class wants to register itself as a listener to a Game object it must inherit the Game::EventHandler class. On state change events a corresponing method is called on the EventHandler interface of each listener. This is what the code looks like:
class Game
{
public:
class EventHandler
{
public:
EventHandler();
virtual ~EventHandler();
virtual void onGameStateChanged(Game * inGame) = 0;
virtual void onLinesCleared(Game * inGame, int inLineCount) = 0;
private:
EventHandler(const EventHandler&);
EventHandler& operator=(const EventHandler&);
};
static void RegisterEventHandler(ThreadSafe<Game> inGame, EventHandler * inEventHandler);
static void UnregisterEventHandler(ThreadSafe<Game> inGame, EventHandler * inEventHandler);
typedef std::set<EventHandler*> EventHandlers;
EventHandlers mEventHandlers;
private:
typedef std::set<Game*> Instances;
static Instances sInstances;
};
void Game::RegisterEventHandler(ThreadSafe<Game> inGame, EventHandler * inEventHandler)
{
ScopedReaderAndWriter<Game> rwgame(inGame);
Game * game(rwgame.get());
if (sInstances.find(game) == sInstances.end())
{
LogWarning("Game::RegisterEventHandler: This game object does not exist!");
return;
}
game->mEventHandlers.insert(inEventHandler);
}
void Game::UnregisterEventHandler(ThreadSafe<Game> inGame, EventHandler * inEventHandler)
{
ScopedReaderAndWriter<Game> rwgame(inGame);
Game * game(rwgame.get());
if (sInstances.find(game) == sInstances.end())
{
LogWarning("Game::UnregisterEventHandler: The game object no longer exists!");
return;
}
game->mEventHandlers.erase(inEventHandler);
}
There are two problems that I often experience with this kind of pattern:
A listener object wants to unregister itself on an already deleted object resulting in a crash.
A event is fired to a listener that no longer exists. This happens most often in multithreaded code. Here's a typical scenario:
The game state changes in a worker thread. We want the notification to occur in the main thread.
The event is wrapped in a boost::function and sent as a PostMessage to the main thread.
A short time later this function object is processed by the main thread while the Game object is already deleted. The result is a crash.
My current workaround is the one that you can see in above code sample. I made the UnregisterEventHandler a static method which checks against a list of instances. This does help, but I find it a somewhat hackish solution.
Does anyone know of a set of guidelines on how to cleanly and safely implement notifier/listener system? Any advice on how to avoid the above pitfalls?
PS: If you need more information in order to answer this question you can find the relevant code online here: Game.h, Game.cpp, SimpleGame.h, SimpleGame.cpp, MainWindow.cpp.

Answer 1

The rule of thumb is that delete and new for an object should be near each other. E.g. in constructor and destructor or before and after a call where you use the object. So it's a bad practice to delete an object in another object when the latter one didn't create the former one.
I don't understand how you pack the events. It seems that you have to check whether the game is still alive before processing an event. Or you can use shared_ptr in events and other places to be sure that games are deleted last.

Answer 2

I write a lot of C++ code and needed to create an Observer for some game components I was working on. I needed something to distribute "start of frame", "user input", etc., as events in the game to interested parties. I had the same problem to consider...the firing of an event would lead to the destruction of another observer which may also subsequently fire. I need to handle this. I did not need to handle thread safety, but the design requirement I usually shoot for is to build something simple enough (API wise) that I can put in a few mutexes in the right place and the rest should take care of itself.
I also wanted it to be straight C++, not dependent on the platform or a specific technology (such as boost, Qt, etc.) because I often build and re-use components (and the ideas behind them) across different projects.
Here is the rough sketch of what I came up with as a solution:
The Observer is a singleton with keys (enumerated values, not strings) for Subjects to register interest in. Because it is a singleton, it always exists.
Each subject is derived from a common base class. The base class has an abstract virtual function Notify(...) which must be implemented in derived classes, and a destructor that removes it from the Observer (which it can always reach) when it is deleted.
Inside the Observer itself, if Detach(...) is called while a Notify(...) is in progress, any detached Subjects end up on a list.
When Notify(...) is called on the Observer, it creates a temporary copy of the Subject list. As it iterates over it, it compare it to the recently detached. If the target is not on it, Notify(...) is called on the target. Otherwise, it is skipped.
Notify(...) in the Observer also keeps track of the depth to handle cascading calls (A notifies B, C, D, and the D.Notify(...) triggers a Notify(...) call to E, etc.)
This is what the interface ended up looking like:
/*
The Notifier is a singleton implementation of the Subject/Observer design
pattern. Any class/instance which wishes to participate as an observer
of an event can derive from the Notified base class and register itself
with the Notiifer for enumerated events.
Notifier derived classes MUST implement the notify function, which has
a prototype of:
void Notify(const NOTIFIED_EVENT_TYPE_T& event)
This is a data object passed from the Notifier class. The structure
passed has a void* in it. There is no illusion of type safety here
and it is the responsibility of the user to ensure it is cast properly.
In most cases, it will be "NULL".
Classes derived from Notified do not need to deregister (though it may
be a good idea to do so) as the base class destrctor will attempt to
remove itself from the Notifier system automatically.
The event type is an enumeration and not a string as it is in many
"generic" notification systems. In practical use, this is for a closed
application where the messages will be known at compile time. This allows
us to increase the speed of the delivery by NOT having a
dictionary keyed lookup mechanism. Some loss of generality is implied
by this.
This class/system is NOT thread safe, but could be made so with some
mutex wrappers. It is safe to call Attach/Detach as a consequence
of calling Notify(...).
*/
class Notified;
class Notifier : public SingletonDynamic<Notifier>
{
public:
typedef enum
{
NE_MIN = 0,
NE_DEBUG_BUTTON_PRESSED = NE_MIN,
NE_DEBUG_LINE_DRAW_ADD_LINE_PIXELS,
NE_DEBUG_TOGGLE_VISIBILITY,
NE_DEBUG_MESSAGE,
NE_RESET_DRAW_CYCLE,
NE_VIEWPORT_CHANGED,
NE_MAX,
} NOTIFIED_EVENT_TYPE_T;
private:
typedef vector<NOTIFIED_EVENT_TYPE_T> NOTIFIED_EVENT_TYPE_VECTOR_T;
typedef map<Notified*,NOTIFIED_EVENT_TYPE_VECTOR_T> NOTIFIED_MAP_T;
typedef map<Notified*,NOTIFIED_EVENT_TYPE_VECTOR_T>::iterator NOTIFIED_MAP_ITER_T;
typedef vector<Notified*> NOTIFIED_VECTOR_T;
typedef vector<NOTIFIED_VECTOR_T> NOTIFIED_VECTOR_VECTOR_T;
NOTIFIED_MAP_T _notifiedMap;
NOTIFIED_VECTOR_VECTOR_T _notifiedVector;
NOTIFIED_MAP_ITER_T _mapIter;
// This vector keeps a temporary list of observers that have completely
// detached since the current "Notify(...)" operation began. This is
// to handle the problem where a Notified instance has called Detach(...)
// because of a Notify(...) call. The removed instance could be a dead
// pointer, so don't try to talk to it.
vector<Notified*> _detached;
int32 _notifyDepth;
void RemoveEvent(NOTIFIED_EVENT_TYPE_VECTOR_T& orgEventTypes, NOTIFIED_EVENT_TYPE_T eventType);
void RemoveNotified(NOTIFIED_VECTOR_T& orgNotified, Notified* observer);
public:
virtual void Reset();
virtual bool Init() { Reset(); return true; }
virtual void Shutdown() { Reset(); }
void Attach(Notified* observer, NOTIFIED_EVENT_TYPE_T eventType);
// Detach for a specific event
void Detach(Notified* observer, NOTIFIED_EVENT_TYPE_T eventType);
// Detach for ALL events
void Detach(Notified* observer);
/* The design of this interface is very specific. I could
* create a class to hold all the event data and then the
* method would just have take that object. But then I would
* have to search for every place in the code that created an
* object to be used and make sure it updated the passed in
* object when a member is added to it. This way, a break
* occurs at compile time that must be addressed.
*/
void Notify(NOTIFIED_EVENT_TYPE_T, const void* eventData = NULL);
/* Used for CPPUnit. Could create a Mock...maybe...but this seems
* like it will get the job done with minimal fuss. For now.
*/
// Return all events that this object is registered for.
vector<NOTIFIED_EVENT_TYPE_T> GetEvents(Notified* observer);
// Return all objects registered for this event.
vector<Notified*> GetNotified(NOTIFIED_EVENT_TYPE_T event);
};
/* This is the base class for anything that can receive notifications.
*/
class Notified
{
public:
virtual void Notify(Notifier::NOTIFIED_EVENT_TYPE_T eventType, const void* eventData) = 0;
virtual ~Notified();
};
typedef Notifier::NOTIFIED_EVENT_TYPE_T NOTIFIED_EVENT_TYPE_T;
NOTE: The Notified class has a single function, Notify(...) here. Because the void* is not type safe, I created other versions where notify looks like:
virtual void Notify(Notifier::NOTIFIED_EVENT_TYPE_T eventType, int value);
virtual void Notify(Notifier::NOTIFIED_EVENT_TYPE_T eventType, const string& str);
Corresponding Notify(...) methods were added to the Notifier itself. All these used a single function to get the "target list" then called the appropriate function on the targets. This works well and keeps the receiver from having to do ugly casts.
This seems to work well. The solution is posted on the web here along with the source code. This is a relatively new design, so any feedback is greatly appreciated.