GUI event blocking

JonB · wrote on 2 Mar 2018, 10:13

The GUI code I have has standard connection of button-clicked to handling code. This might be from a "window" (QWidget) or from a dialog (QDialog).

For the first time, I have clicked a button which executes long-running code, doing goodness knows what. It does not seem to do any UI stuff itself, though (if relevant) it may do things like access a database through Qt.

To my "horror", I find that while running I can click the button again and set off the same code, or indeed interact with any other UI element, with all the consequences.

Perhaps because of my background elsewhere, I am used the UI being "blocked"/"non-re-entrant" while executing code in response to UI event like a button click, only accepting new input events when the handling code has run to completion.

If this is not the case with Qt, it sounds like I am responsible for writing code to cause every UI event to "block" any other possible UI event, assuming I want that behaviour. Is that indeed correct? I haven't seen examples doing this, and it would seem to be a very common requirement....

If the answer is that this re-entrant behaviour is a consequence of something the large code is doing --- perhaps it spins an event loop somewhere --- let me know.

sierdzio · 2 Mar 2018, 11:51

Normally in such situation your GUI will be blocked, so this has something to do with the code in that slot your are calling.

See if the slot is not connected by Qt::QueuedConnection by any chance. Or perhaps the slot is using some async operations (like using QNetworkAccessManager, QProcess), then methods return immediately and results are communicated via signals.

Or see if the code calls processEvents() anywhere in that slot. Or maybe it fires off some QThread.

There are many possibilities here, without more info I'm just guessing. But definitely, by default, GUI would freeze if slot connected to a button is doing something CPU intensive on the same thread.

mrjj · 2 Mar 2018, 16:12

Hi
You can use the Qt::BlockingQueuedConnection
as not to allow overlapping runs for same signal.
Dialogs can be used for modal input blocking the whole parent.
The enable property can be used to control other elements for non modal designs.

JonB · 3 Mar 2018, 02:15

@sierdzio

Normally in such situation your GUI will be blocked, so this has something to do with the code in that slot your are calling.

Ah, that is indeed what I needed to know. So the normal behaviour is indeed "block other UI events till slot handler returns in response to original signal", and in that respect Qt behaves to what i am used to from elsewhere.

I have no idea what the mass of code it calls is doing in detail. I don't think the code explicitly has any of the Qt code you mention, but I don't know for sure, and it's going to be hard to track down.

As I mentioned, one thing it will be doing is database access, potentially via QSql... calls. Do these count for any kind of processEvents() or QThread in themselves?.

@mrjj

You can use the Qt::BlockingQueuedConnection
as not to allow overlapping runs for same signal.
Dialogs can be used for modal input blocking the whole parent.
The enable property can be used to control other elements for non modal designs.

Hmm. "Same signal" won't do me: it might be invoked by QPushButton::clicked, but potentially user could do any kind of other interaction which invokes a slot, I can't rely on "clicked" only. Dialogs may block parent, but that doesn't help, it's other inputs on same dialog I would be worried about. Unless that's what you mean. It turns out the "window" is actually a stacked tab widget thing which has buttons, table view/widgets etc. Are you saying I can set enabled property of that (the stacked widget) to false to stop any other input controls responding?

kshegunov · M mrjj 2 Mar 2018, 11:11

@mrjj said in GUI event blocking:

You can use the Qt::BlockingQueuedConnection

If this is not directed to a separate thread (i.e. the receiver object isn't outside the main thread), you'd deadlock the GUI thread, so use at your own discretion!

ambershark · 3 Mar 2018, 02:16

@JonB Yea it sounds like you are doing a lot in a single GUI click. You should probably have some indication to the user that processing is occurring. Database access, etc, can take a while. In this case maybe pop up a progress bar or message or something.

The GUI should be locked while that single click is being processed unless it's multithreaded. For example if your click handler does something like this:

void SomeWidget::onClick()
{
   // do something
   // do something else
   // done
}

No events on the GUI should be processed until onClick exits. So once onClick returns it could be clicked again even if it is technically still processing or waiting on a database or something. So if you are using any sort of asynchronous call or a separate thread it should definitely lock the GUI.

You could also use a flag to denote processing and if that button is clicked a second time ignore the click. However that only protects that single button and not any action in the GUI.

JonB · 3 Mar 2018, 07:56

@ambershark , and others

I now understand that "normal" Qt processing should be "block-until-return-from-click-handler". (Which is a relief to my general handling of other aspects of UI!)

However, in the case of the code (large, spaghetti, and completely uncommented) I am looking at, here is what I experience:

Click button to set off large "computation" of some kind. It's going to take, say, 20 seconds to complete.
Wait, say, a couple of seconds for it to "get going". (Don't know if this is required; there's obviously some delay between clicks anyway.)
Click that button again, or any other button/widget on the "stacked widget page".
Code immediately starts executing in response to this new UI action. Verified in debugger and/or messages.

The question for me is: what might it be in the original code processing which is allowing the second UI interaction to proceed? I need to track it down. To the best of my knowledge the code does not explicitly do things like create threads or process an event loop. It does, however, do database querying, which is why I keep asking if that's enough to cause the re-entrancy?

I know all about progress bars and blocking flags. Indeed, with a blocking flag, the blocking flag immediately gets hit in #4, proving the point, but not robust enough as you say, e.g. does not cover clicking any other widget in UI.

The code is so monolithic & interdependent that, say, commenting out sections to see if behaviour altered is difficult/impossible. So I don't know how to approach debugging to discover what this massive block of unintelligible code is doing which allows this re-entrancy, so that I can deal with it properly.

Can anyone suggest a technique (Python debugging, I'm afraid, not native C++) to aid me in tracking down where the re-entrancy is coming from?

kshegunov · 3 Mar 2018, 08:44

@JonB said in GUI event blocking:

Can anyone suggest a technique (Python debugging, I'm afraid, not native C++) to aid me in tracking down where the re-entrancy is coming from?

Look for a thread being started in the debugger whenever that piece of code's executed, if that's so then you have your answer.

JonB · 3 Mar 2018, 08:45

@kshegunov
Yeah, thanks, but my knowledge of the PyCharm debugger is insufficient to "Look for a thread being started in the debugger"! Not your problem, but mine :)

What about stack traceback when the second click causes "re-entrance" to the function, which I can break on? I'm wondering whether that might show me enough to get back, presumably through an event processing loop invocation, to where the code is allowing that to happen?

kshegunov · 3 Mar 2018, 08:50

@JonB said in GUI event blocking:

What about stack traceback when the second click causes "re-entrance" to the function, which I can break on? I'm wondering whether that might show me enough to get back, presumably through an event processing loop invocation, to where the code is allowing that to happen?

Yes, it should be possible, but I imagine it'd land you in the slot handling the actual button click signal. I'd start investigating from that slot anyway - there should be some hint in it. Probably a class is starting a thread internally or something like this. The other option is that someone is calling QCoreAppliction::processEvents intermittently to process the pending events, as already mentioned by @sierdzio, then you will actually need to find the code doing it and act accordingly. My best advice is to get into step-by-step debugging and inspect what's called where in that signal handler, it may be quite a lot of work, but I don't know of any quick and easy way.

As a "workaround", you could disable the button after it being clicked, and reenable it whenever the long operation has completed, so you don't get the problem with calling the heavy code twice.

JonB · 3 Mar 2018, 08:54

@kshegunov
Thanks, I will investigate.

One more time, as nobody has said "yay" or "nay", do you think making one of the QSql... calls to access the database spins a thread/event loop which could be the cause???

As a "workaround", you could disable the button after it being clicked, and reenable it whenever the long operation has completed, so you don't get the problem with calling the heavy code twice.

As I said, I have already done this. Better than nothing. But not robust enough, as there are many other widgets on the window which user could interact with... And even is they do not re-invoke the long-running code, I cannot allow any other code to execute while one piece is already executing, for all the obvious reasons...

kshegunov · 3 Mar 2018, 09:33

@JonB said in GUI event blocking:

do you think making one of the QSql... calls to access the database spins a thread/event loop which could be the cause???

Nope, I don't. As far as I have worked and looked at the SQL classes none of them (drivers) does use threading and/or event loops.

I cannot allow any other code to execute while one piece is already executing, for all the obvious reasons...

Then another option is to "quick-fix" it by setting a boolean flag just before the long operation starts, if it's raised then just quick return and do nothing, otherwise raise it and do the crunching. Whenever the operation's done, just lower the flag.

JonB · 3 Mar 2018, 09:38

@kshegunov

Thanks for the heads-up on QSql calls not being the cause.

I cannot allow any other code to execute while one piece is already executing, for all the obvious reasons...

Then another option is to "quick-fix" it by setting a boolean flag just before the long operation starts, if it's raised then just quick return and do nothing, otherwise raise it and do the crunching. Whenever the operation's done, just lower the flag.

That's what I'm saying I've done. But I think (politely) you're missing the point on re-entrancy. You're concentrating only on preventing the long-runner from being re-invoked. Yes, I do that. But you need to think the other way round as well. While something is running, no other action is safe to perform. You mustn't, for example, close the dialog, or even request something from the database, or anything else you can think of, in response to a new UI action (let's pretend the window has 100 other buttons/views/widgets/tooltips/menus which do goodness knows what; I don't want to go through 100 functions testing for flag), because we don't know what state the application, or the database, or anything else, is in, until long-runner completes. Which is what we would get if clicking to run long-runner in the first place blocked, as it's "supposed" to.... And that's what I want to get back to.

kshegunov · 3 Mar 2018, 09:40

@JonB said in GUI event blocking:

But I think (politely) you're missing the point on re-entrancy.

Probably, but then again I don't have the code at hand, so I'm just more or less guessing. Your point seems warranted though, so the only thing I can add here is: have a light and easy debugging. :)

JonB · 3 Mar 2018, 09:41

@kshegunov
Thanks; as I said, my comment was intended "politely". I have tried to clarify the situation in previous post. Once a single UI operation is underway, it's not safe to allow any other one to proceed, unless you've written code in a very particular way. And I didn't even write this code, let alone know just what it does!

ambershark · 3 Mar 2018, 19:43

@JonB said in GUI event blocking:

The question for me is: what might it be in the original code processing which is allowing the second UI interaction to proceed?

Well there are really only 2 things that can do this. Threading and events. So either the signal handler for the click is exiting after sending events or it's exiting after signalling a thread.

To be honest it should be pretty easy to debug. Just point a breakpoint in the signal handler for that click event, then step through it. Step over function calls and test how long they take. If everything exits quickly (what I'm guessing will happen) and the signal handler exits, before your data is processed, then you just start stepping in to each of the functions in that click handler until you see what they are doing.

At some point you will find one that signals something to do some work and exits (via thread or event). That is your problem. Address that area and you should be good to go. I would just use something easy like a modal progress dialog that locks up the gui automatically until you signal it to quit when you get back from the processing of the "click".

Oh and it always sucks working on someone else's code.. Especially confusing/bad code. When I get in messy code I use a tool called SourceTrail to navigate code.

https://www.sourcetrail.com

JonB · 3 Jun 2018, 10:39

@ambershark
OK, now had a chance to break and examine stack.

Nobody said this code would be easy...

The "long running operation" involves producing HTML, and saving to PDF file (think of it like a report). I do this via QWebEngineView (partly because this is in shared code, at other times the generated HTML is displayed to the user and is editable prior to producing the PDF). In this case the view is never displayed interactively to the user, but it still follows that route.

Because QWebEngineView is asynchronous (its own thread, I believe), and here we need the HTML/PDF synchronously, before we can save the PDF to file we must await QWebEngineView finished loading (else we get empty PDF saved to file). I have code there like:

    def synchronousRenderHtml(self):
        # see synchronousWebViewLoaded() below
        self.rendered = False
        self.webView.loadFinished.connect(self.synchronousWebViewLoaded)
        #
        # see http://doc.qt.io/qt-5/qtwebenginewidgets-qtwebkitportingguide.html
        # because setHtml() is asynchronous, the call to make editable has to be moved into synchronousWebViewLoaded()
        htmlFile = filesystemfunctions.findFileInDataOrCodeDirectory(Paths.s166Html())
        self.webView.setHtml(self.html, QUrl.fromLocalFile(str(htmlFile)))
        #
        # see synchronousWebViewLoaded() below
        if not self.rendered:
            self.renderLoop.exec()

    def synchronousWebViewLoaded(self):
        # This function is called on self.webView.loadFinished() from synchronousRenderHtml() above
        # The code uses this because self.webView.setHtml() is asynchronous
        # so in effect this implements a "blocking" call (with event processing) in synchronousRenderHtml()
        # This is the only place in code where a QEventLoop is used explicitly
        # It's all very complicated, and I did wish this behaviour was not needed
        # but it seems, at minimum in the case where this dialog is used non-interactively, it is :(
        self.qWebEngineView.page().runJavaScript("document.documentElement.contentEditable = true")
        self.rendered = True
        # cause the self.renderLoop.exec() in synchronousRenderHtml() above to exit now
        self.renderLoop.quit()

Turns out, the re-entrant call for the top-level button being pressed a second time is coming from the self.renderLoop.exec() (which is awaiting the self.renderLoop.quit()).

Cutting a long story short, unless someone can suggest a better way than this self.renderLoop.exec() principle(?), I think I have to stick with that. Hence the button re-entrancy, and I'll have to live with the need to using a re-entrancy blocking flag....

kshegunov · 3 Jul 2018, 09:09

@JonB said in GUI event blocking:

Cutting a long story short, unless someone can suggest a better way than this self.renderLoop.exec() principle(?),

Not in this particular case. But I have a suggestion:
Grab the mouse and keyboard in synchronousRenderHtml and discard those events in your event() override while self.rendered is false and you're owning the user input. In synchronousWebViewLoaded release back the mouse and keyboard. Tread carefully though, as grabbing the mouse and keyboard can make your system unusable (i.e. use -nograb with gdb to prevent you from cursing). ;)

JonB · 3 Jul 2018, 08:42

@kshegunov
I did think about that "dirty" way of handling things.

However (apart from the debugging issue, I use Python + PyCharm debugger, not gdb, and I wouldn't hold your breath that it will have a -nograb functionality!), the problem will be which "event() override" you have in mind.

Because, as I said, the QWebEngineView, which produces the HTML/PDF asynchronously and has the synchronousRenderHtml() call, is not displayed to the user in this case, the UI events are being received to the original dialog where a button is pressed to produce the output. This in itself has no connection to the ("invisible") QWebEngineView --- it doesn't even know one is being used, the implementation is invisible to it. I would have to modify its event handling. And since that would break encapsulation, I really don't feel it would be a good idea to implement!

So I will soldier on with the "blocking flag" I currently have to prevent "re-entrancy", which at least works reasonably well in practice. At least I now understand why this whole behaviour is caused.

kshegunov · 7 Mar 2018, 09:21

@JonB said in GUI event blocking:

the problem will be which "event() override" you have in mind.

That would be the event() override of the class (widget) in which you grab the input. When you grab the input events all of them are redirected to the widget that called the mentioned functions, so you'd need to filter them in that same class. Basically, you grab the input in the widget/dialog/w/e and filter the input there. Whenever the operation's done (i.e. you connect the finished signal to a slot in that same dialog you release the mouse/keyboard.