Why it uses d->eventFilters.prepend(obj) not append(obj) in function(QObject::installEventFilter),i want to know why design it in such way.I just curious about it.
void QObject::installEventFilter(QObject *obj)
{
Q_D(QObject);
if (!obj)
return;
if (d->threadData != obj->d_func()->threadData) {
qWarning("QObject::installEventFilter(): Cannot filter events for objects in a different thread.");
return;
}
// clean up unused items in the list
d->eventFilters.removeAll((QObject*)0);
d->eventFilters.removeAll(obj);
d->eventFilters.prepend(obj);
}
It's done that way because the most recently installed event filter is to be processed first, i.e. it needs to be at the beginning of the filter list. The filters are invoked by traversing the list in sequential order from begin() to end().
The most recently installed filter is to be processed first because the only two simple choices are to either process it first or last. And the second choice is not useful: when you filter events, you want to decide what happens before anyone else does. Well, but then some new user's filter will go before yours, so how that can be? As follows: event filters are used to amend functionality - functionality that already exists. If you added a filter somewhere inside the existing functionality, you'd effectively be interfacing to a partially defined system, with unknown behavior. After all, even Qt's implementation uses event filters. They provide the documented behavior. By inserting your event filter last, you couldn't be sure at all what events it will see - it'd all depend on implementation details of every layer of functionality above your filter.
A system with some event filter installed is like a layer of skin on the onion - the user of that system only sees the skin, not what's inside, not the implementation. But they can add their own skin on top if they wish so, and implement new functionality that way. They can't dig into the onion, because they don't know what's in it. Of course that's a generalization: they don't know because it doesn't form an API, a contract between them and the implementation of the system. They are free to read the source code and/or reverse engineer the system, and then insert the event filter anywhere in the list they wish. After all, once you get access to QObjectPrivate, you can modify the event filter list as you wish. But then you're responsible for the behavior of not only what you added on top of the public API, but of many of the underlying layers too - and your responsibility broadens. Updating the toolkit becomes next to impossible, because you'd have to audit the code and/or verify test coverage to make sure that something somewhere in the internals didn't get broken.
Related
I recently refactored half my app to use the provider pattern and i now have a problem. The main issue is that i need to initialise controllers in the init (e.g. a text controller to have an initial value or the list size of a tab controller)
How am i meant to init Controllers if the data i need has to come from the state in the build method.
For example.
// This must go in the build as it requires state
myTabsController = TabController(length: myState.list.length, vsync: this);
I'm initialisng the controller every time it builds now... How am i meant to put this in the init but still access the state variables (as there is no context).
I've tried using the afterFirstLayout() callback from the AfterLayoutMixin library but that just causes more problems. Currently with the tab bar it flashes error as no tab initialised for the first frame and then displays properly when the afterFirstLayout is called and initialises the tab. This seems like a hacky fix
I would like to learn more about how to use this pattern properly and what would be the best solution to this problem.
Feel free to ask me to clarify more.
Thanks for your help.
I have what I believe is a very common scenario... I'm building a dashboard of components that will be driven by some datasource. At the top of the view would be a series of filters (e.g. a date range). When the date range is updated, the components on the screen would need to update their data based on the selected range. This would in turn force the individual components that are slave to that picker to need to fetch new data (async action/XHR) based on the newly selected range.
There can be many components on the screen and the user may wish to add/remove available displays, so it is not as simple as always refreshing the data for all components because they may or may not be present.
One way I thought to handle this was in the action dispatched when a new date range is selected was to figure out what components are on screen (derived from the Store) and dispatch async actions to fetch the data for those components. This seems like a lot of work will go into the DATE_CHANGED action.
Another alternative might be to detect date range changes in store.subscribe() callbacks from each of the components. This seems to decouple the logic to fetch the data from the action that caused this to happen. However, I thought it was bad practice (or even an error) to dispatch while dispatching. Sure I can wrap it in a setTimeout, but that feels wrong too.
Third thing that came to mind was just doing fetch calls directly in the component's store.subscribe() and dispatching when those return, but I thought this breaks the connect model.
This seems like a common pattern to fetch based on state changes, but I don't know where its best to put those. Any good documentation / examples on the above problem?
Don't use store.subscribe for this. When DATE_CHANGED reaches the reducer it's meant for, simply change the application state (I'm assuming the date range is part of the store somehow). So you have something like state.rangeStart and state.rangeEnd.
You didn't mention what view rendering library you're using, so I can only describe how this is typically done with React:
The components know wether they are currently mounted (visible) or not, so redux doesn't need to be concerned with that. What you need is a way to detect that state.rangeStart or state.rangeEnd changed.
In React there is a lifecycle hook for that (componentWillReceiveProps or getDerivedStateFromProps in the newest release). In this handler you dispatch async redux actions that fetch the data the component needs. Your view library will probably have something similar.
The components display some kind of "empty" or "loading" state while you're waiting for the new data typically. So a good practice is to invalidate/clear data from the store in the reducer that handles the DATE_CHANGED action. For example, if state.listOfThings (an array) entirely depends on the date range, you would set it to an empty array as soon as the date changes: return { ...state, listOfThings: [] }. This causes the components to display that data is being fetched again.
When all the async redux actions went through the REQUEST -> SUCCESS/FAILURE cycle and have populated the store with the data, connected components will automatically render it. This is kind of its own chapter, look into redux async actions if you need more information.
The tricky part are interdependencies between the components and the application they're rendering. If two different dashboard components for example want to fetch and render state.listOfThings for the current date range, you don't want to fetch this data twice. So there needs to be a way to detected that 1) the data range has changed but also 2) a request to fetch listOfThings is already on its way. This is usually done with boolean flags in the state: state.isFetchingListOfThings. The async actions fetching this data cause the reducer to set this flag to true. Your components need to be aware of this and dispatch actions conditionally: if (props.rangeStart !== nextProps.rangeStart && !nextProps.isFetchingListOfThings) { props.fetchListOfThings(); }.
I want to know how to ensure the sort indicator in the horizontal header of a QTableView is updated when a programmatic sort is performed on the model.
Here's the problem:
QStandardItemModel model(3,1);
QTableView view;
view.setModel( &model );
// Populate the model ensuring it is not in a sorted order
for( int row = 0; row < model.rowCount(); ++row )
{
model.setItem( row , 0 ,
new QStandardItem(QString::number((row+1)%model.rowCount())));
}
view.setSortingEnabled( true );
// At this point everything is consistent since enabling the sorting
// triggers a sort that matches the indicator in the horizontalHeader (see A)
model.sort( 0 , Qt::AscendingOrder );
// However at this point the sort order has been reversed but the
// header's sort indicator remains unchanged (see B)
A: B:
As you can see the sort indicator remains the same and therefore is inconsistent with the actual sort order.
In my application I have two views that interact with the same model and sorting can be triggered from either of them. I don't see anything in QAbstractItemModel that signals when a sort has been performed. It seems like QHeaderView/TableView assume that they are the only thing that can trigger a sort.
Does Qt provide facilities for coping with this that I'm missing? If not, what's the best way of keeping the sort indicator up-to-date without breaking the encapsulation of the multiple views on the model too much?
One of the ItemDataRole enumerators available since Qt 4.8 is InitialSortOrderRole.
http://qt-project.org/doc/qt-4.8/qt.html#ItemDataRole-enum
It should therefore be possible to transmit sort order information through the QAbstractItemModel::headerData method.
I've tried this however and found that QTableView and QHeaderView do not seem to update in response to changes in this headerData role. A customised header view would appear to be necessary...
It might be worth it because passing this information via the model allows any number of views to synchronise without any external agent having to track all the views in existence so that it can distribute notifications. It would also work seamlessly through model proxy stacks such as those built with QSortFilterModelProxy.
The solution I've come up with to avoid breaking encapsulation too much is
to have a signal on each view (on QTableView the sortIndicatorChanged signal suffices and on my custom view I have added a similar signal).
the manager of views connects to these signals
when any view emits such a signal the manager of views calls a slot on all the other views so that they can synchronise their sort indicators
I still feel like I might be missing something - surely this is a common problem? It seems to me that QAbstractItemModel should have a way of transmitting sort-order information to views...
In my main window (QMainWindow) I have a QTableView (named commandsTableView). Now I want to react on its selection changes.
I made a slot and connected it manually to ui.commandsTableView->selectionModel(). All works fine.
But then I thought: why not use auto-connection (especially that there will be more connections to be done)? At least it will add more force to consistent naming rules.
Yet I wasn't able to find proper name syntax. I tried:
on_commandsTableView_selectionModel_selectionChanged,
on_commandsTableViewSelectionModel_selectionChanged,
on_commandsTableView_selectionChanged,
on_commandsTableView___selectionChanged
but neither worked. In all cases there is there is a message on output when running the app (with corresponding slot name, here only first given as an example):
QMetaObject::connectSlotsByName: No matching signal for on_commandsTableView_selectionModel_selectionChanged(QItemSelection,QItemSelection)
(Why there are no assertions in response for connection errors - that I cannot understand. I lost much time wondering what is wrong before I spotted those - and alike - messages on output.)
The object returned by ui.commandsTableView->selectionModel() has an empty name. But setting it to selectionModel prior to making a call to connectSlotsByName doesn't help either.
According to the documentation connectSlotsByName() only supports signatures like
void on_<object name>_<signal name>(<signal parameters>);
According to the sources that's the only form it checks (watch how it collects a list of children, then matches parent's method names against names of the children).
Hence, to be able to use auto-connection you would have needed a named selection model, which would continue existing from the call to connectSlotsByName() onwards. Each time you change the selection model (not likely) or the model (likely) you'd have to name the selection model and auto-connect again. But alas connectSlotsByName() will duplicate all other connections as it doesn't seem to check if connections are unique, so we have to connect signals to such dynamic children as models, scenes etc manually.
I think it's
on_selectionModel_selectionChanged(const QItemSelection & selected, const QItemSelection & deselected)
I am trying to get my Windows State Machine workflow to communicate with end users. The general pattern I am trying to implement within a StateActivity is:
StateInitializationActivity: Send a message to user requesting an answer to a question (e.g. "Do you approve this document?"), together with the context for...
...EventDrivenActivity: Deal with answer sent by user
StateFinalizationActivity: Cancel message (e.g. document is withdrawn and no longer needs approval)
This all works fine if the StateActivity is a "Leaf State" (i.e. has no child states). However, it does not work if I want to use recursive composition of states. For non-leaf states, StateInitialization and StateFinalization do not run (I confirmed this behaviour by using Reflector to inspect the StateActivity source code). The EventDrivenActivity is still listening, but the end user doesn't know what's going on.
For StateInitialization, I thought that one way to work around this would be to replace it with an EventDrivenActivity and a zero-delay timer. I'm stuck with what to do about StateFinalization.
So - does anyone have any ideas about how to get a State Finalization Activity to always run, even for non-leaf states?
Its unfortunate that the structure of "nested states" is one of a "parent" containing "children", the designer UI re-enforces this concept. Hence its quite natural and intuative to think the way you are thinking. Its unfortunate because its wrong.
The true relationship is one of "General" -> "Specific". Its in effect a hierachical class structure. Consider a much more familar such relationship:-
public class MySuperClass
{
public MySuperClass(object parameter) { }
protected void DoSomething() { }
}
public class MySubClass : MySuperClass
{
protected void DoSomethingElse() { }
}
Here MySubClass inherits DoSomething from SuperClass. The above though is broken because the SuperClass doesn't have a default constructor. Also parameterised constructor of SuperClass is not inherited by SubClass. In fact logically a sub-class never inherits the constructors (or destructors) of the super-class. (Yes there is some magic wiring up default constructors but thats more sugar than substance).
Similarly the relationship between StateAcivities contained with another StateActivity is actually that the contained activity is a specialisation of the container. Each contained activity inherits the set of event driven activities of the container. However, each contained StateActivity is a first class discrete state in the workflow same as any other state.
The containing activity actual becomes an abstract, it can not be transitioned to and importantly there is no real concept of transition to a state "inside" another state. By extension then there is no concept of leaving such an outer state either. As a result there is no initialization or finalization of the containing StateActivity.
A quirk of the designer allows you to add a StateInitialization and StateFinalization then add StateActivities to a state. If you try it the other way round the designer won't let you because it knows the Initialization and Finalization will never be run.
I realise this doesn't actually answer your question and I'm loath to say in this case "It can't be done" but if it can it will be a little hacky.
OK, so here’s what I decided to do in the end. I created a custom tracking service which looks for activity events corresponding to entering or leaving the states which are involved in communication with end users. This service enters decisions for the user into a database when the state is entered and removes them when the state is left. The user can query the database to see what decisions the workflow is waiting on. The workflow listens for user responses using a ReceiveActivity in an EventDrivenActivity. This also works for decisions in parent ‘superstates’. This might not be exactly what a "Tracking Service" is meant to be for, but it seems to work
I've thought of another way of solving the problem. Originally, I had in mind that for communications I would use the WCF-integrated SendActivity and ReceiveActivity provided in WF 3.5.
However, in the end I came to the conclusion that it's easier to ignore these activities and implement your own IEventActivity with a local service. IEventActivity.Subscribe can be used to indicate to users that there is a question for them to answer and IEventActivity.Unsubscribe can be used to cancel the question. This means that separate activities in the State's inialization and finalization blocks are not required. The message routing is done manually using workflow queues and the user's response is added to the queue with appropriate name. I used Guid's for the queue names, and these are passed to the user during the IEventActivity.Subscribe call.
I used the 'File System Watcher' example in MSDN to work out how to do this.
I also found this article very insructive: http://www.infoq.com/articles/lublinksy-workqueue-mgr