The two previous chapters discussed the details of GLib's Dynamic Type System and its signal control system. The GObject library also contains an implementation for a base fundamental type named GObject. GObject is a fundamental classed instantiable type. It implements: Memory management with reference counting Construction/Destruction of instances Generic per-object properties with set/get function pairs Easy use of signals All the GNOME libraries which use the GLib type system (like GTK+ and GStreamer) inherit from GObject which is why it is important to understand the details of how it works. The g_object_new family of functions can be used to instantiate any GType which inherits from the GObject base type. All these functions make sure the class and instance structures have been correctly initialized by GLib's type system and then invoke at one point or another the constructor class method which is used to: Allocate and clear memory through g_type_create_instance, Initialize the object's instance with the construction properties. Although one can expect all class and instance members (except the fields pointing to the parents) to be set to zero, some consider it good practice to explicitly set them. Objects which inherit from GObject are allowed to override this constructor class method: they should however chain to their parent constructor method before doing so: GObject *(* constructor) (GType gtype, guint n_properties, GObjectConstructParam *properties); The example below shows how MamanBar overrides the parent's constructor: #define MAMAN_TYPE_BAR (maman_bar_get_type ()) #define MAMAN_BAR(obj) (G_TYPE_CHECK_INSTANCE_CAST ((obj), MAMAN_TYPE_BAR, MamanBar)) #define MAMAN_IS_BAR(obj) (G_TYPE_CHECK_INSTANCE_TYPE ((obj), MAMAN_TYPE_BAR)) #define MAMAN_BAR_CLASS(klass) (G_TYPE_CHECK_CLASS_CAST ((klass), MAMAN_TYPE_BAR, MamanBarClass)) #define MAMAN_IS_BAR_CLASS(klass) (G_TYPE_CHECK_CLASS_TYPE ((klass), MAMAN_TYPE_BAR)) #define MAMAN_BAR_GET_CLASS(obj) (G_TYPE_INSTANCE_GET_CLASS ((obj), MAMAN_TYPE_BAR, MamanBarClass)) typedef struct _MamanBar MamanBar; typedef struct _MamanBarClass MamanBarClass; struct _MamanBar { GObject parent_instance; /* instance members */ }; struct _MamanBarClass { GObjectClass parent_class; /* class members */ }; /* will create maman_bar_get_type and set maman_bar_parent_class */ G_DEFINE_TYPE (MamanBar, maman_bar, G_TYPE_OBJECT); static GObject * maman_bar_constructor (GType gtype, guint n_properties, GObjectConstructParam *properties) { GObject *obj; { /* Always chain up to the parent constructor */ obj = G_OBJECT_CLASS (maman_bar_parent_class)->constructor (gtype, n_properties, properties); } /* update the object state depending on constructor properties */ return obj; } static void maman_bar_class_init (MamanBarClass *klass) { GObjectClass *gobject_class = G_OBJECT_CLASS (klass); gobject_class->constructor = maman_bar_constructor; } static void maman_bar_init (MamanBar *self) { /* initialize the object */ } If the user instantiates an object MamanBar with: MamanBar *bar = g_object_new (MAMAN_TYPE_BAR, NULL); If this is the first instantiation of such an object, the maman_bar_class_init function will be invoked after any maman_bar_base_class_init function. This will make sure the class structure of this new object is correctly initialized. Here, maman_bar_class_init is expected to override the object's class methods and setup the class' own methods. In the example above, the constructor method is the only overridden method: it is set to maman_bar_constructor. Once g_object_new has obtained a reference to an initialized class structure, it invokes its constructor method to create an instance of the new object. Since it has just been overridden by maman_bar_class_init to maman_bar_constructor, the latter is called and, because it was implemented correctly, it chains up to its parent's constructor. In order to find the parent class and chain up to the parent class constructor, we can use the maman_bar_parent_class pointer that has been set up for us by the G_DEFINE_TYPE macro. Finally, at one point or another, g_object_constructor is invoked by the last constructor in the chain. This function allocates the object's instance' buffer through g_type_create_instance which means that the instance_init function is invoked at this point if one was registered. After instance_init returns, the object is fully initialized and should be ready to answer any user-request. When g_type_create_instance returns, g_object_constructor sets the construction properties (i.e. the properties which were given to g_object_new) and returns to the user's constructor which is then allowed to do useful instance initialization... The process described above might seem a bit complicated, but it can be summarized easily by the table below which lists the functions invoked by g_object_new and their order of invocation: Invocation time Function Invoked Function's parameters Remark First call to g_object_new for target type target type's base_init function On the inheritance tree of classes from fundamental type to target type. base_init is invoked once for each class structure. I have no real idea on how this can be used. If you have a good real-life example of how a class' base_init can be used, please, let me know. target type's class_init function On target type's class structure Here, you should make sure to initialize or override class methods (that is, assign to each class' method its function pointer) and create the signals and the properties associated to your object. interface' base_init function On interface' vtable interface' interface_init function On interface' vtable Each call to g_object_new for target type target type's class constructor method: GObjectClass->constructor On object's instance If you need to complete the object initialization after all the construction properties are set, override the constructor method and make sure to chain up to the object's parent class before doing your own initialization. In doubt, do not override the constructor method. type's instance_init function On the inheritance tree of classes from fundamental type to target type. the instance_init provided for each type is invoked once for each instance structure. Provide an instance_init function to initialize your object before its construction properties are set. This is the preferred way to initialize a GObject instance. This function is equivalent to C++ constructors. Readers should feel concerned about one little twist in the order in which functions are invoked: while, technically, the class' constructor method is called before the GType's instance_init function (since g_type_create_instance which calls instance_init is called by g_object_constructor which is the top-level class constructor method and to which users are expected to chain to), the user's code which runs in a user-provided constructor will always run after GType's instance_init function since the user-provided constructor must (you've been warned) chain up before doing anything useful. The memory-management API for GObjects is a bit complicated but the idea behind it is pretty simple: the goal is to provide a flexible model based on reference counting which can be integrated in applications which use or require different memory management models (such as garbage collection, aso...). The methods which are used to manipulate this reference count are described below. /* Refcounting */ gpointer g_object_ref (gpointer object); void g_object_unref (gpointer object); /* * Weak References */ typedef void (*GWeakNotify) (gpointer data, GObject *where_the_object_was); void g_object_weak_ref (GObject *object, GWeakNotify notify, gpointer data); void g_object_weak_unref (GObject *object, GWeakNotify notify, gpointer data); void g_object_add_weak_pointer (GObject *object, gpointer *weak_pointer_location); void g_object_remove_weak_pointer (GObject *object, gpointer *weak_pointer_location); /* * Cycle handling */ void g_object_run_dispose (GObject *object); The functions g_object_ref/g_object_unref respectively increase and decrease the reference count.These functions are thread-safe as of GLib 2.8. The reference count is, unsurprisingly, initialized to one by g_object_new which means that the caller is currently the sole owner of the newly-created reference. When the reference count reaches zero, that is, when g_object_unref is called by the last client holding a reference to the object, the dispose and the finalize class methods are invoked. Finally, after finalize is invoked, g_type_free_instance is called to free the object instance. Depending on the memory allocation policy decided when the type was registered (through one of the g_type_register_* functions), the object's instance memory will be freed or returned to the object pool for this type. Once the object has been freed, if it was the last instance of the type, the type's class will be destroyed as described in and . The table below summarizes the destruction process of a GObject: Invocation time Function Invoked Function's parameters Remark Last call to g_object_unref for an instance of target type target type's dispose class function GObject instance When dispose ends, the object should not hold any reference to any other member object. The object is also expected to be able to answer client method invocations (with possibly an error code but no memory violation) until finalize is executed. dispose can be executed more than once. dispose should chain up to its parent implementation just before returning to the caller. target type's finalize class function GObject instance Finalize is expected to complete the destruction process initiated by dispose. It should complete the object's destruction. finalize will be executed only once. finalize should chain up to its parent implementation just before returning to the caller. The reason why the destruction process is split is two different phases is explained in . Last call to g_object_unref for the last instance of target type interface' interface_finalize function On interface' vtable Never used in practice. Unlikely you will need it. interface' base_finalize function On interface' vtable Never used in practice. Unlikely you will need it. target type's class_finalize function On target type's class structure Never used in practice. Unlikely you will need it. type's base_finalize function On the inheritance tree of classes from fundamental type to target type. base_init is invoked once for each class structure. Never used in practice. Unlikely you will need it. Weak References are used to monitor object finalization: g_object_weak_ref adds a monitoring callback which does not hold a reference to the object but which is invoked when the object runs its dispose method. As such, each weak ref can be invoked more than once upon object finalization (since dispose can run more than once during object finalization). g_object_weak_unref can be used to remove a monitoring callback from the object. Weak References are also used to implement g_object_add_weak_pointer and g_object_remove_weak_pointer. These functions add a weak reference to the object they are applied to which makes sure to nullify the pointer given by the user when object is finalized. Note: the following section was inspired by James Henstridge. I guess this means that all praise and all curses will be directly forwarded to him. GObject's memory management model was designed to be easily integrated in existing code using garbage collection. This is why the destruction process is split in two phases: the first phase, executed in the dispose handler is supposed to release all references to other member objects. The second phase, executed by the finalize handler is supposed to complete the object's destruction process. Object methods should be able to run without program error (that is, without segfault :) in-between the two phases. This two-step destruction process is very useful to break reference counting cycles. While the detection of the cycles is up to the external code, once the cycles have been detected, the external code can invoke g_object_dispose which will indeed break any existing cycles since it will run the dispose handler associated to the object and thus release all references to other objects. Attentive readers might now have understood one of the rules about the dispose handler we stated a bit sooner: the dispose handler can be invoked multiple times. Let's say we have a reference count cycle: object A references B which itself references object A. Let's say we have detected the cycle and we want to destroy the two objects. One way to do this would be to invoke g_object_dispose on one of the objects. If object A releases all its references to all objects, this means it releases its reference to object B. If object B was not owned by anyone else, this is its last reference count which means this last unref runs B's dispose handler which, in turn, releases B's reference on object A. If this is A's last reference count, this last unref runs A's dispose handler which is running for the second time before A's finalize handler is invoked ! The above example, which might seem a bit contrived can really happen if your GObject's are being handled by language bindings. I would thus suggest the rules stated above for object destruction are closely followed. Otherwise, Bad Bad Things will happen. One of GObject's nice features is its generic get/set mechanism for object properties. When an object is instantiated, the object's class_init handler should be used to register the object's properties with g_object_class_install_property (implemented in gobject.c). The best way to understand how object properties work is by looking at a real example on how it is used: /************************************************/ /* Implementation */ /************************************************/ enum { PROP_0, PROP_MAMAN_NAME, PROP_PAPA_NUMBER }; static void maman_bar_set_property (GObject *object, guint property_id, const GValue *value, GParamSpec *pspec) { MamanBar *self = MAMAN_BAR (object); switch (property_id) { case PROP_MAMAN_NAME: g_free (self->priv->name); self->priv->name = g_value_dup_string (value); g_print ("maman: %s\n", self->priv->name); break; case PROP_PAPA_NUMBER: self->priv->papa_number = g_value_get_uchar (value); g_print ("papa: %u\n", self->priv->papa_number); break; default: /* We don't have any other property... */ G_OBJECT_WARN_INVALID_PROPERTY_ID (object, property_id, pspec); break; } } static void maman_bar_get_property (GObject *object, guint property_id, GValue *value, GParamSpec *pspec) { MamanBar *self = MAMAN_BAR (object); switch (property_id) { case PROP_MAMAN_NAME: g_value_set_string (value, self->priv->name); break; case PROP_PAPA_NUMBER: g_value_set_uchar (value, self->priv->papa_number); break; default: /* We don't have any other property... */ G_OBJECT_WARN_INVALID_PROPERTY_ID (object, property_id, pspec); break; } } static void maman_bar_class_init (MamanBarClass *klass) { GObjectClass *gobject_class = G_OBJECT_CLASS (klass); GParamSpec *pspec; gobject_class->set_property = maman_bar_set_property; gobject_class->get_property = maman_bar_get_property; pspec = g_param_spec_string ("maman-name", "Maman construct prop", "Set maman's name", "no-name-set" /* default value */, G_PARAM_CONSTRUCT_ONLY | G_PARAM_READWRITE); g_object_class_install_property (gobject_class, PROP_MAMAN_NAME, pspec); pspec = g_param_spec_uchar ("papa-number", "Number of current Papa", "Set/Get papa's number", 0 /* minimum value */, 10 /* maximum value */, 2 /* default value */, G_PARAM_READWRITE); g_object_class_install_property (gobject_class, PROP_PAPA_NUMBER, pspec); } /************************************************/ /* Use */ /************************************************/ GObject *bar; GValue val = { 0, }; bar = g_object_new (MAMAN_TYPE_SUBBAR, NULL); g_value_init (&val, G_TYPE_CHAR); g_value_set_char (&val, 11); g_object_set_property (G_OBJECT (bar), "papa-number", &val); g_value_unset (&val); The client code just above looks simple but a lot of things happen under the hood: g_object_set_property first ensures a property with this name was registered in bar's class_init handler. If so, it calls object_set_property which first walks the class hierarchy, from bottom, most derived type, to top, fundamental type to find the class which registered that property. It then tries to convert the user-provided GValue into a GValue whose type is that of the associated property. If the user provides a signed char GValue, as is shown here, and if the object's property was registered as an unsigned int, g_value_transform will try to transform the input signed char into an unsigned int. Of course, the success of the transformation depends on the availability of the required transform function. In practice, there will almost always be a transformation Its behaviour might not be what you expect but it is up to you to actually avoid relying on these transformations. which matches and conversion will be carried out if needed. After transformation, the GValue is validated by g_param_value_validate which makes sure the user's data stored in the GValue matches the characteristics specified by the property's GParamSpec. Here, the GParamSpec we provided in class_init has a validation function which makes sure that the GValue contains a value which respects the minimum and maximum bounds of the GParamSpec. In the example above, the client's GValue does not respect these constraints (it is set to 11, while the maximum is 10). As such, the g_object_set_property function will return with an error. If the user's GValue had been set to a valid value, g_object_set_property would have proceeded with calling the object's set_property class method. Here, since our implementation of Foo did override this method, the code path would jump to foo_set_property after having retrieved from the GParamSpec the param_id It should be noted that the param_id used here need only to uniquely identify each GParamSpec within the FooClass such that the switch used in the set and get methods actually works. Of course, this locally-unique integer is purely an optimization: it would have been possible to use a set of if (strcmp (a, b) == 0) {} else if (strcmp (a, b) == 0) {} statements. which had been stored by g_object_class_install_property. Once the property has been set by the object's set_property class method, the code path returns to g_object_set_property which calls g_object_notify_queue_thaw. This function makes sure that the "notify" signal is emitted on the object's instance with the changed property as parameter unless notifications were frozen by g_object_freeze_notify. g_object_thaw_notify can be used to re-enable notification of property modifications through the "notify" signal. It is important to remember that even if properties are changed while property change notification is frozen, the "notify" signal will be emitted once for each of these changed properties as soon as the property change notification is thawed: no property change is lost for the "notify" signal. Signal can only be delayed by the notification freezing mechanism. It sounds like a tedious task to set up GValues every time when one wants to modify a property. In practice one will rarely do this. The functions g_object_set_property and g_object_get_property are meant to be used by language bindings. For application there is an easier way and that is described next. It is interesting to note that the g_object_set and g_object_set_valist (vararg version) functions can be used to set multiple properties at once. The client code shown above can then be re-written as: MamanBar *foo; foo = /* */; g_object_set (G_OBJECT (foo), "papa-number", 2, "maman-name", "test", NULL); This saves us from managing the GValues that we were needing to handle when using g_object_set_property. The code above will trigger one notify signal emission for each property modified. Of course, the _get versions are also available: g_object_get and g_object_get_valist (vararg version) can be used to get numerous properties at once. These high level functions have one drawback - they don't provide a return result. One should pay attention to the argument types and ranges when using them. A known source of errors is to e.g. pass a gfloat instead of a gdouble and thus shifting all subsequent parameters by four bytes. Also forgetting the terminating NULL will lead to unexpected behaviour. Really attentive readers now understand how g_object_new, g_object_newv and g_object_new_valist work: they parse the user-provided variable number of parameters and invoke g_object_set on the parameters only after the object has been successfully constructed. Of course, the "notify" signal will be emitted for each property set.