How many event queues are there in Qt application? Events like click etc are enqueued in the queue. And you can also enque events using function like postEvent. Are all these events in the same queue? if yes is there a way to define different queues?
Normally, there is at most one event queue per each thread.
It is possible to have additional event queues, with two implementation strategies that differ on whether you remain compatible with QCoreApplication::postEvent.
If you don't care about QCoreApplication::postEvent working with your queue, everything is up to you. You can implement the queue in whatever way you wish. Note that Qt doesn't implement a way of marking a QObject as belonging to a particular event queue. When you're ready to deliver the events from your queue, you simply invoke QCoreApplication::notify, passing it the target object pointer, and event pointer. Easy-peasy. You don't have any control over the events that are delivered to any and all objects from the default queue, of course.
Another way is to remain compatible with QCoreApplication::postEvent. Namely, you somehow "mark" an object as having its events handled by your queue. You intercept the event about to be delivered to the target, enqueue it, and handle it yourself as needed.
This is the QStateMachine-like hack. It is good. And bad. And things in between. It's worth knowing how it's done and why.
Hierarchical state machines typically need to control the delivery of events and inject their own events into the queue, in front of other events. This is so that the state transition events are delivered in right order in relation to the events that cause the transitions; furthermore sometimes a transition-causing event might need to be delivered multiple times, retained for later delivery, etc.
This is all in the face of the rigid event lifetime enforced by the default event dispatching system. To work around it, QStateMachine implements its own event queue. Whenever you declare a transition on an event that would be delivered to some object, the state machine installs itself as an event filter on that object.
Then, whenever the original event reaches the target object, the filter intercepts the event and prevents its delivery to the target object. Now it must make a copy of the event, and insert it into its own queue. The copy must be made, because the event will be deleted by the event dispatcher as soon as the control leaves the event filter and/or the target object's event() method.
Unfortunately, before Qt 6, QEvents were not cloneable - at least not publicly so. There was some clone functionality hidden in Qt's innards, usable by user code, but it was a private API, and not based on the virtual copy constructor idiom.
Qt 6 has added the QEvent::clone method, and events should be presumed cloneable. Custom events in legacy code that wasn’t ported to Qt 6 won’t support this, and their clones will not be fully functional if they carry data.
Using the filter-based idiom/hack, you can implement a notion of a QObject belonging to a certain event queue. When your queue filters the events on the target object, clones them, and enqueues them for separate delivery, the object functionally resides on your event queue. When no such filtering takes place, the object resides on the default per-thread queue. You can also only intercept events of the type(s) your queue is interested in, just like the QStateMachine does.
Related
I am developing an application with Qt whereby I want to update my GUI from another unrelated class with functions executing in a different thread.
I registered a callback function from the Qt GUI class to the other class using standard C++ constructs std::function and std::bind.
On trying to execute the GUI callback function from the other class, my program crashes with "qt cannot send events to objects owned by a different thread".
My questions are, what is happening here exactly. Why is it not permissive to communicate between two threads like this. Is it possible to resolve this in some way , or does Qt just not allow other unrelated functions to update the GUI in this way ?
For information , I am using portaudio libraries on windows. It executes a callback in a different thread. When I receive audio, I try to update my GUI.
Thank you in advance
You're probably being thrown off by the "send events" phrase. In Qt, this phrase has a strict technical meaning. Namely:
to send events means to use QCoreApplication::sendEvent, which then immediately invokes QObject::event on the receiver. It is an error to send events when the sending and receiving threads are different.
to post events means to use QCoreApplication::postEvent, which posts the event to the receiving object thread's event queue. The QObject::event on the receiver is invoked later by the receiver thread's event loop.
When objects live in different threads, you're not allowed to send events, only to post them. This points to the underlying cause of your issue/
Any method you access on a QWidget or a derived class must be accessed on its thread(), i.e. the main thread. Everything inside of QWidget can use sendEvent freely as the sending thread and receiving object's thread are identical by contract. By calling QWidget methods from the wrong thread, you break that contract by indirectly using sendEvent and thus fail.
Only methods that you yourself have implemented and are thread-safe can be called from other threads - and they can't use sendEvent.
Does Qt offer functionality to know the number of queued signals that are pending for a given slot to process? Is there a way to clear them? For example, if several emits are done on a signal connected to a given slot, how could someone know the amount of these emitted signals?
QMetaObject::Connection class has a laconic interface and does not seem to offer related functionality. Deleting the object that receives the signals, hence destroying the connections, solves the problem. But is there a way to do this without disconnecting the slots or deleting the receiving objects?
The reason you ask the question most likely indicates that your design is broken. The signals and slots are a mechanism to decouple code. The objects connected together should behave themselves no matter how many senders or receiver are there, and certainly shouldn't attempt to track such matters!
It'd be wiser to fix the issue at the source by altering the design. If you are uffering from an event storm e.g. due to changing a widget's data in a slot, the slot should be very lightweight and only schedule an update of the widget by calling update(), but never forcing an immediate repaint. This leverages repaint event compression done by Qt. You might wish to compress your own events too.
Connection types in Qt
Signals and slots in Qt can be delivered using a direct, queued or blocking connection. The automatic type is not really a fixed connection type. It is an instruction to resolve the type for every receiver, at every signal emission into either a direct or queued type.
The direct connection is like any indirect function call: nothing is queued, and the slot is called from within the body of the signal method:
// all direct-connected slots/functors are invoked before mySignal() returns
emit mySignal();
The queued connection posts a QMetaCallEvent to the receiving object thread's event loop. That event contains the arguments of the call, or carries the functor. Is is handled by QObject::event(). You can certainly intercept such events. For details, see this question.
As far as I know, it is not possible to access the queue.
First of all, if the slot is in a QWidget subclass, in the GUI thread, then you can just update the member variables and call update(), and then just use the current values in paintEvent() when it gets called. These are automatically compressed, so there will be just one repaint event, no matter how many times update() gets called.
However, if the slot is not related to painting, or not in the GUI thread at all, then you need something else.
A simple solution to many cases needing this is to use a 2nd slot and a single-shot QTimer with delay 0 (or even longer delay if desirable).
Here's some example code, which should give you the idea of what I mean:
// in constructor, set mActualSlotTimer to
// singleshot, interval 0, parent this (needed for multi-threaded use)
// and connect timeout() to privateActualSlot()
// public slot for receiving the signal from outside
void MyClass::actualSlot(int data) {
// class member to store the new data value until it can be set
mNewData = data;
// restart the timer, no matter if it was already running or not
mActualSlotTimer.start();
}
// "private" slot for actually doing the change
void MyClass::privateActualSlot() {
// maybe useful: if (this->mData == this->mNewData) return;
mData = mNewData;
// do whatever else needs to be done!
}
Obviously, if your public slot does not actually take any parameters, then you don't need mData and mNewData.
One thing to note about this approach is, it works on all connections, it's not limited to just Qt::QueuedConnecton. As a consequence, it also makes using Qt::BlockingQueuedConnection kinda pointless.
A disclaimery note: I briefly checked Qt source code and it seems using timer with interval 0 should be ok: restarting the timer will work as expected. But if there still seems to be too many calls to privateActualSlot, then providing a suitable interval might be necessary. I have usually wanted a little delay (5 ms for example) to throttle things down a bit more than "as often as possible", so have not tested this extensively with interval 0.
During debugging, I want to see what's awaits my program's event loop.
It's probably flooded, and I want to see by what signals, without (manually) adding specific log-message to every Q_EMIT.
Possible solutions might be watching some internal-qt-data structure that contains the events-queue (Is there such thing? how?)
Or -
Write a log message for every signal emitted (Is that possible?).
Any other ideas?
(QT 4.8 on Windows, using visual studio 2012)
Signals and events are two things that don't have anything to do with each other.
I want to see what's awaits my program's event loop. It's probably flooded.
First of all, let's get the nomenclature straight:
an event queue is where events are stored until delivery;
an event loop is what drains the event queue and delivers events to QObjects,
an event flood happens when, on average, during the delivery of each event there is more than one event added to the queue.
There are two reasons only why an event queue might get flooded:
It takes too long to process some events (e.g. when your code blocks): the drain rate of the queue is lower than the fill rate due to timing.
You're adding more than one event per each event delivered (on average): the fill rate of the queue is higher than the drain rate due to event multiplication - this is completely unrelated to any timing. An apt name for it would be an event storm.
To detect code the blocks for too long, you can use a tool I wrote for another answer.
To know how many events are waiting for any given thread, use the undocumented qGlobalPostedEventsCount(). You add that to the code of the tool linked-to above.
Not sure if this is sufficient for you but you can try installing event filters in between QObjects that implement eventFilter() like this:
class MyWidget : public QWidget
{
QGraphicsView myView;
MyWidget()
{
myView->installEventFilter(this);
// incoming events to myView are shown on the standard output
}
};
You can get more creative with this reading the docs.
The Qt documentation for Events and Filters states:
It is also possible to filter all events for the entire application, by installing an event filter on the QApplication or QCoreApplication object. Such global event filters are called before the object-specific filters. This is very powerful, but it also slows down event delivery of every single event in the entire application.
Therefore, you can create an event filter on the QApplication or QCoreApplication and monitor all events, checking their type.
Alternatively, QCoreApplication uses the virtual notify function to deliver events to objects. Overriding QCoreApplication would allow you to see both the event and QObject to which the event will initially* be delivered.
*Note that events are propagated to parent objects, if the receiving object ignores the event.
If you choose to use notify, be aware of the future direction for this function:
Future direction: This function will not be called for objects that live outside the main thread in Qt 6. Applications that need that functionality should find other solutions for their event inspection needs in the meantime. The change may be extended to the main thread, causing this function to be deprecated.
I am learning Qt framework (C++) and was wondering if QT has any mechanisms to protect slots from being called before an object is fully initialized.
Consider Class A constructor:
A::A() {
mTreeView = new QTreeView();
...
connect(mTreeView, &QTreeView::customContextMenuRequested,
this, &A::OnContextMenuRequested);
...
}
My worry is the user would be able to right-click on the tree-view before A's constructor has finished. Another context is as follows:
A::A() {
mQObj = new MyQObject();
connect(mQObj, &MyQObject::SomeEvent, this, &A::OnEvent);
}
A::InitB() { mB = new B(); }
A::OnEvent() { mB.doSomething(); }
Here, the doSomething() method can be called before InitB() runs.
Do I have to worry about such scenarios? And if so, is there a way to avoid these issues without having to initialize all your objects first, then going back and connecting the events afterwards separately?
You don't have to worry about such scenarios in most cases, because events are delivered in the same thread. There's no "hidden multithreading" going on you have to care about. If you don't explicitly call a function in the constructor of A that causes events to be processed, you're safe and your current method, slot etc.'s execution is completed before the next event is processed.
That said, the cases where a new event is processed and thus other code (event handlers, slots) is executed, are:
the execution of the current event handler, slot etc. is finished (the code where you create A) and Qt returns to the event loop to wait for the next event. In your case, that's after the A instance is fully constructed.
you start a local event loop (Create a QEventLoop object and call exec())
you call QCoreApplication::processEvents()
you call exec() on a QDialog
1) is the normal mode Qt operates in: You start app.exec(), which starts the event loop. Everything after that is directly or indirectly triggered by an event (user input, timer, I/O). Event happens and is added to the event loop's event queue. The event loop calls the event handlers for the event. When the event handlers are completed, the event loop picks the next event and calls the handlers for it.
So everything happens in an orderly fashion, one event after another, unless one of the event handlers (say, a slot listening to a button's clicked() signal) does one of 2, 3 or 4. Then Qt processes the next event in-place, .i.e. where exec() or processEvents() is called. The event handlers/slots are executed accordingly. Then exec()/processEvents() returns. Now all the certainties one would have without calling exec()/processEvents() are gone, unfortunately: The user could have done random stuff, things could be arbitrarily changed or deleted (even the this pointer, if the user closed the window, for example). Thus especially 2) and 3) are error-prone and usually lead to major headaches, so I would avoid them whenever possible, or at least be aware of the potential issues.
Now there's the case that you use multithreading yourself. As all the actual UI and related user events is handled in a single main thread, multithreading here usually means that you have threads doing non-UI work and that interacts with your UI thread by calling functions on objects living in the UI thread, modifying data shared by both threads, or using cross-thread signal/slot connections.
If you don't use signal/slots but direct method calls from the secondary thread to the UI thread or modify shared data, the usual multithreading rules apply: Things will go majorly wrong unless you know what you're doing and use synchronization accordingly. Qt's UI classes aren't thread-safe.
If you use signal/slots, the calls are /queued/, i.e. if thread B emits a signal and you receive it in a slot in your main thread, the call of the slot is delivered in the same manner as a user event: the same rules as for user events apply. Unless you do one of 2, 3, 4, the slot won't be called before your event handler/slot returns. So cross-thread signal/slots connections are a way of /message passing/ where the messages are delivered via the event loop.
In the Qt world, what is the difference of events and signal/slots?
Does one replace the other? Are events an abstraction of signal/slots?
In Qt, signals and events are both implementations of the Observer pattern. They are used in different situations because they have different strengths and weaknesses.
First of all let's define what we mean by 'Qt event' exactly: a virtual function in a Qt class, which you're expected to reimplement in a base class of yours if you want to handle the event. It's related to the Template Method pattern.
Note how I used the word "handle". Indeed, here's a basic difference between the intent of signals and events:
You "handle" events
You "get notified of" signal emissions
The difference is that when you "handle" the event, you take on the responsibility to "respond" with a behavior that is useful outside the class. For example, consider an app that has a button with a number on it. The app needs to let the user focus the button and change the number by pressing the "up" and "down" keyboard keys. Otherwise the button should function like a normal QPushButton (it can be clicked, etc). In Qt this is done by creating your own little reusable "component" (subclass of QPushButton), which reimplements QWidget::keyPressEvent. Pseudocode:
class NumericButton extends QPushButton
private void addToNumber(int value):
// ...
reimplement base.keyPressEvent(QKeyEvent event):
if(event.key == up)
this.addToNumber(1)
else if(event.key == down)
this.addToNumber(-1)
else
base.keyPressEvent(event)
See? This code presents a new abstraction: a widget that acts like a button, but with some extra functionality. We added this functionality very conveniently:
Since we reimplemented a virtual, our implementation automatically became encapsulated in our class. If Qt's designers had made keyPressEvent a signal, we would need to decide whether to inherit QPushButton or just externally connect to the signal. But that would be stupid, since in Qt you're always expected to inherit when writing a widget with a custom behavior (for good reason - reusability/modularity). So by making keyPressEvent an event, they convey their intent that keyPressEvent is just a basic building block of functionality. If it were a signal, it'd look like a user-facing thing, when it's not intended to be.
Since the base-class-implementation of the function is available, we easily implement the Chain-of-responsibility pattern by handling our special cases (up&down keys) and leaving the rest to the base class. You can see this would be nearly impossible if keyPressEvent were a signal.
The design of Qt is well thought out - they made us fall into the pit of success by making it easy to do the right thing and hard to do the wrong thing (by making keyPressEvent an event).
On the other hand, consider the simplest usage of QPushButton - just instantiating it and getting notified when it's clicked:
button = new QPushButton(this)
connect(button, SIGNAL(clicked()), SLOT(sayHello())
This is clearly meant to be done by the user of the class:
if we had to subclass QPushButton every time we want some button to notify us of a click, that would require a lot of subclasses for no good reason! A widget that always shows a "Hello world" messagebox when clicked is useful only in a single case - so it's totally not reusable. Again, we have no choice but to do the right thing - by connecting to it externally.
we may want to connect several slots to clicked() - or connect several signals to sayHello(). With signals there is no fuss. With subclassing you would have to sit down and ponder some class diagrams until you decide on an appropriate design.
Note that one of the places QPushButton emits clicked() is in its mousePressEvent() implementation. That doesn't mean clicked() and mousePressEvent() are interchangable - just that they're related.
So signals and events have different purposes (but are related in that both let you "subscribe" to a notification of something happening).
I don’t like the answers so far. – Let me concentrate on this part of the question:
Are events an abstraction of signal/slots?
Short answer: no. The long answer raises a “better” question: How are signals and events related?
An idle main loop (Qt’s for example) is usually “stuck” in a select() call of the operating system. That call makes the application “sleep”, while it passes a bunch of sockets or files or whatever to the kernel asking for: if something changes on these, let the select() call return. – And the kernel, as the master of the world, knows when that happens.
The result of that select() call could be: new data on the socket connect to X11, a packet to a UDP port we listen on came in, etc. – That stuff is neither a Qt signal, nor a Qt event, and the Qt main loop decides itself if it turns the fresh data into the one, the other or ignores it.
Qt could call a method (or several) like keyPressEvent(), effectively turning it into a Qt event. Or Qt emits a signal, which in effect looks up all functions registered for that signal, and calls them one after the other.
One difference of those two concepts is visible here: a slot has no vote on whether other slots registered to that signal will get called or not. – Events are more like a chain, and the event handler decides if it interrupts that chain or not. Signals look like a star or tree in this respect.
An event can trigger or be entirely turned into a signal (just emit one, and don’t call “super()”). A signal can be turned into an event (call an event handler).
What abstracts what depends on the case: the clicked()-signal abstracts mouse events (a button goes down and up again without too much moving around). Keyboard events are abstractions from lower levels (things like 果 or é are several key strokes on my system).
Maybe the focusInEvent() is an example of the opposite: it could use (and thus abstract) the clicked() signal, but I don’t know if it actually does.
The Qt documentation probably explains it best:
In Qt, events are objects, derived
from the abstract QEvent class, that
represent things that have happened
either within an application or as a
result of outside activity that the
application needs to know about.
Events can be received and handled by
any instance of a QObject subclass,
but they are especially relevant to
widgets. This document describes how
events are delivered and handled in a
typical application.
So events and signal/slots are two parallel mechanisms accomplishing the same things. In general, an event will be generated by an outside entity (for example, keyboard or mouse wheel) and will be delivered through the event loop in QApplication. In general, unless you set up the code, you will not be generating events. You might filter them through QObject::installEventFilter() or handle events in subclassed object by overriding the appropriate functions.
Signals and Slots are much easier to generate and receive and you can connect any two QObject subclasses. They are handled through the Metaclass (have a look at your moc_classname.cpp file for more), but most of the interclass communication that you will produce will probably use signals and slots. Signals can get delivered immediately or deferred via a queue (if you are using threads).
A signal can be generated.
Events are dispatched by the event loop. Each GUI program needs an event loop, whatever you write it Windows or Linux, using Qt, Win32 or any other GUI library. As well each thread has its own event loop. In Qt "GUI Event Loop" (which is the main loop of all Qt applications) is hidden, but you start it calling:
QApplication a(argc, argv);
return a.exec();
Messages OS and other applications send to your program are dispatched as events.
Signals and slots are Qt mechanisms. In the process of compilations using moc (meta-object compiler), they are changed to callback functions.
Event should have one receiver, which should dispatch it. No one else should get that event.
All slots connected to the emitted signal will be executed.
You shouldn't think of Signals as events, because as you can read in the Qt documentation:
When a signal is emitted, the slots connected to it are usually executed immediately, just like a normal function call. When this happens, the signals and
slots mechanism is totally independent
of any GUI event loop.
When you send an event, it must wait for some time until the event loop dispatches all events that came earlier. Because of this, execution of the code after sending event or signal is different. Code following sending an event will be run immediately. With the signals and slots mechanisms it depends on the connection type. Normally it will be executed after all slots. Using Qt::QueuedConnection, it will be executed immediately, just like events. Check all connection types in the Qt documentation.
There is an article that discusses event processing in some detail: http://www.packtpub.com/article/events-and-signals
It discussions the difference between events and signals here:
Events and signals are two parallel mechanisms used to accomplish the
same thing. As a general difference, signals are useful when using a
widget, whereas events are useful when implementing the widget. For
example, when we are using a widget like QPushButton, we are more
interested in its clicked() signal than in the low-level mouse press
or key press events that caused the signal to be emitted. But if we
are implementing the QPushButton class, we are more interested in the
implementation of code for mouse and key events. Also, we usually
handle events but get notified by signal emissions.
This seems to be a common way of talking about it, as the accepted answer uses some of the same phrases.
Note, please see helpful comments below on this answer from Kuba Ober, that make me wonder if it might be a bit simplistic.
TL;DR: Signals and slots are indirect method calls. Events are data structures. So they are quite different animals.
The only time when they come together is when slot calls are made across thread boundaries. The slot call arguments are packed up in a data structure and get sent as an event to the receiving thread's event queue. In the receiving thread, the QObject::event method unpacks the arguments, executes the call, and possibly returns the result if it was a blocking connection.
If we're willing to generalize to oblivion, one could think of events as as a way of invoking the target object's event method. This is an indirect method call, after a fashion - but I don't think it's a helpful way of thinking about it, even if it's a true statement.
'Event processing' by Leow Wee Kheng says:
Jasmine Blanchette says:
The main reason why you would use events rather than standard function calls, or signals and slots, is that events can be used both synchronously and asynchronously (depending on whether you call sendEvent() or postEvents()), whereas calling a function or invoking a slot is always synchronous. Another advantage of events is that they can be filtered.
Events (in a general sense of user/network interaction) are typically handled in Qt with signals/slots, but signals/slots can do plenty of other things.
QEvent and its subclasses are basically just little standardized data packages for the framework to communicate with your code. If you want to pay attention to the mouse in some way, you only have to look at the QMouseEvent API, and the library designers don't have to reinvent the wheel every time you need to figure out what the mouse did in some corner of the Qt API.
It is true that if you're waiting for events (again in the general case) of some sort, your slot will almost certainly accept a QEvent subclass as an argument.
With that said, signals and slots can certainly be used without QEvents, although you'll find that the original impetus for activating a signal will often be some kind of user interaction or other asynchronous activity. Sometimes, however, your code will just reach a point where firing off a certain signal will be the right thing to do. For example, firing off a signal connected to a progress bar during a long process doesn't involve a QEvent up to that point.
Another minor pragmatic consideration: emitting or receiving signals requires inheriting QObject whereas an object of any inheritance can post or send an event (since you invoke QCoreApplication.sendEvent() or postEvent()) This is usually not an issue but: to use signals PyQt strangely requires QObject to be the first super class, and you might not want to rearrange your inheritance order just to be able to send signals.)
In my opinion events are completely redundant and could be thrown out. There is no reason why signals could not be replaced by events or events by signals, except that Qt is already set up as it is. Queued signals are wrapped by events and events could conceivably be wrapped by signals, for example:
connect(this, &MyItem::mouseMove, [this](QMouseEvent*){});
Would replace the convenience mouseMoveEvent() function found in QWidget (but not in QQuickItem anymore) and would handle mouseMove signals that a scene manager would emit for the item. The fact that the signal is emitted on behalf of the item by some outside entity is unimportant and happens quite often in the world of Qt components, even though it is supposedly not allowed (Qt components often circumvent this rule). But Qt is a conglomerate of many different design decisions and pretty much cast in stone for fear of breaking old code (which happens often enough anyway).