Global static QPixmapCache in Qt internals

[kshegunov]

@JKSH said in QApplication in std::thread:

I believe it is fixable with m_qPlotsControl->deleteLater();... but that needs to be done before QApplication::exec() returns, because QDeferredDeleteEvent won't be processed without an event loop.

I still don't understand this. For what reason should that be the case, and why would this generate a crash? I'm pretty sure deferred deletes are processed after the event loop has exited (see for example the QThread::finished -> QObject::deleteLater), but even if that weren't the case the QPushButton's destructor is going to be run before the thread actually finishes and before the application object's destructor's run (there's no difference between main() and any other function from c++'s point of view).

[kshegunov]

@JKSH said in QApplication in std::thread:

The code below prints "Bye" but not "Widget destroyed", proving that the Widget's destructor is not run:

I meant connecting to the aboutToQuit signal, but anyhow it's not that important.

Since the widget creates a timer in the background, then it's probably receiving background events/signals. delete could cause the event/signal handler to act on a dangling pointer -- that's how it could generate a crash.

I think this shouldn't happen, but can't investigate it currently.

Anyway, I was trying to say that @Suthiro can't simply replace delete m_qPlotsControl; with m_qPlotsControl->deleteLater(); in-place at https://stackoverflow.com/questions/67257329/weird-qpushbutton-bug . Some re-shuffling is required.

I suppose, but that timer message should go away, otherwise the code can break at any moment (this applies to your code too - the one you mentioned where you get the warning).

[Chris Kawa]

I reproduced this on Windows and the message about timers comes from a destructor of a static QPixmapCache object that happens on the main() thread. That's not something the user's code does. It's in Qt's internals.

Right, so remember the rule that no QObject should outlive the application object and this means no static QObjects? Apparently Qt itself doesn't respect that rule and also creates static objects that run code in the main thread no matter where the application object is created :/

I would consider this a Qt bug and expect it to crash all over the place.

EDIT: I missed that JKSH already mentioned this, but yeah, this seems to be the culprit.

[kshegunov]

@Chris-Kawa said in QApplication in std::thread:

I reproduced this on Windows and the message about timers comes from a destructor of a static QPixmapCache object that happens on the main() thread. That's not something the user's code does. It's in Qt's internals.

Hm, interesting. Do you have a woboq location?

Right, so remember the rule that no QObject should outlive the application object and this means no static QObjects?
Apparently Qt itself doesn't respect that rule and also creates static objects that run code in the main thread no matter where the application object is created :/

Yeah. Not living up to its own expectations it seems. :(

I would consider this a Qt bug and expect it to crash all over the place.

Me too, which is what I wrote somewhere upstairs - this should be fixed at the vendor side.

[Chris Kawa]

@kshegunov said in QApplication in std::thread:

Hm, interesting. Do you have a woboq location?

Instance declaration is here:
https://code.woboq.org/qt5/qtbase/src/gui/image/qpixmapcache.cpp.html#pm_cache

And here's Windows with MSVC crash callstack on exit:

1   QPMCache::~QPMCache                                                                                   qpixmapcache.cpp 269
2   ``anonymous namespace'::Q_QGS_pm_cache::innerFunction'::`2'::Holder::~Holder                          Qt5Guid             
3   ``anonymous namespace'::Q_QGS_pm_cache::innerFunction'::`2'::`dynamic atexit destructor for 'holder'' Qt5Guid             
4   initterm_e                                                                                            ucrtbased           
5   initterm_e                                                                                            ucrtbased           
6   initterm_e                                                                                            ucrtbased           
7   execute_onexit_table                                                                                  ucrtbased           
8   __scrt_dllmain_uninitialize_c                                                                         utility.cpp      399
9   dllmain_crt_process_detach                                                                            dll_dllmain.cpp  182
10  dllmain_crt_dispatch                                                                                  dll_dllmain.cpp  220
11  dllmain_dispatch                                                                                      dll_dllmain.cpp  293
12  _DllMainCRTStartup                                                                                    dll_dllmain.cpp  335
13  RtlActivateActivationContextUnsafeFast                                                                ntdll               
14  LdrShutdownProcess                                                                                    ntdll               
15  RtlExitUserProcess                                                                                    ntdll               
16  FatalExit                                                                                             KERNEL32            
17  wassert                                                                                               ucrtbased           
18  wassert                                                                                               ucrtbased           
19  exit                                                                                                  ucrtbased           
20  __scrt_common_main_seh                                                                                exe_common.inl   297

I used this simplified code:

int main(int argc, char *argv[])
{
    std::thread thr([](int argc, char *argv[])
    {
        QApplication a(argc, argv);
        QPushButton btn("Push");
        btn.show();
        a.exec();
    }, argc, argv);

    thr.join();
}

[kshegunov]

Yes, this is constructed appropriately, however the deallocation should've been tied to the QCoreApplication::aboutToQuit. I doubt anyone is going to bother fixing it though, I'm not sure if it's not going to be discarded with "this is unsupported" even ...

[JKSH]

Good detective work, @Chris-Kawa!

I encountered this a few years ago but had forgotten most of the details.

@kshegunov said in Global static QPixmapCache in Qt internals:

Yes, this is constructed appropriately, however the deallocation should've been tied to the QCoreApplication::aboutToQuit. I doubt anyone is going to bother fixing it though, I'm not sure if it's not going to be discarded with "this is unsupported" even ...

Since a QPixmap requires a QGuiApplication, would a simple(?) and correct fix be to allocate the QPixmapCache in the QGuiApplication constructor, and deallocate it from the QGuiApplication destructor?

[kshegunov]

Possibly. Or it can be a pseudo-singleton (similarly to QCoreApplication) and be initialized on demand in the instance retrieving function (dropping the global static that it currently uses). You could poke Thiago, but as written, I wouldn't hold my breath.

[Christian Ehrlicher]

There are some reports about such an issue:
https://bugreports.qt.io/browse/QTBUG-48709
https://bugreports.qt.io/browse/QTBUG-21807

And it does not crash here for me on Linux and Windows (MSVC, debug-build) with 5.15.x

[kshegunov]

@Christian-Ehrlicher said in Global static QPixmapCache in Qt internals:

And it does not crash here for me on Linux and Windows (MSVC, debug-build) with 5.15.x

We are talking about the QTimers from another thread message. JKSH, split the threads. The other stuff (the supposed crash) contines to befuddle me.

[Christian Ehrlicher]

I also don't have a warning about the timer with the testcase on windows/msvc :)

[Suthiro]

@kshegunov

We are talking about the QTimers from another thread message. JKSH, split the threads. The other stuff (the supposed crash) contines to befuddle me.

I don't know how exactly the crash in the original thread is connected with this. However, the crash and the message about timers are appearing together. They appear after any interaction with simple QPushButton. Using deleteLater() as suggested by @JKSH solves the crash, but the message persists.

@Christian-Ehrlicher said in Global static QPixmapCache in Qt internals:

And it does not crash here for me on Linux and Windows (MSVC, debug-build) with 5.15.x

To clarify: crash appears only when one tries to unload a library with QApplication inside. The library in my case is loaded by lua.exe, but I doubt it is lua-related problem.

[jeremy_k]

@kshegunov said in Global static QPixmapCache in Qt internals:

Possibly. Or it can be a pseudo-singleton (similarly to QCoreApplication) and be initialized on demand in the instance retrieving function (dropping the global static that it currently uses). You could poke Thiago, but as written, I wouldn't hold my breath.

That sounds like a source and binary compatibility break, unless instance() is used from within the existing static functions.

A thread_local version would be another solution, that could be implemented without changing any user source code. Keep the global static interface object for binary compatibility, and move the actual QObject-based implementation to thread_local storage for the gui thread.

The QPixmapCache documentation already says that the cache is only usable from the "application's main thread".

[kshegunov]

@jeremy_k said in Global static QPixmapCache in Qt internals:

That sounds like a source and binary compatibility break, unless instance() is used from within the existing static functions.

Adding non-virtual functions doesn't brake API or ABI compatibility. Also this is an internal member, any global getter function could do, it's not really necessary to expose it to the user.

A thread_local version would be another solution, that could be implemented without changing any user source code. Keep the global static interface object for binary compatibility, and move the actual QObject-based implementation to thread_local storage for the gui thread.

The QPixmapCache documentation already says that the cache is only usable from the "application's main thread".

It's an internal class, I didn't intend to modify the QPixmapCache to begin with. (see the woboq link in Chris' post)

[jeremy_k]

@kshegunov said in Global static QPixmapCache in Qt internals:

@jeremy_k said in Global static QPixmapCache in Qt internals:

That sounds like a source and binary compatibility break, unless instance() is used from within the existing static functions.

Adding non-virtual functions doesn't brake API or ABI compatibility. Also this is an internal member, any global getter function could do, it's not really necessary to expose it to the user.

I had read it as a suggestion to introduce a static QPixmapCache::instance(), and make the currently static QPixmapCache member functions non-static. I autocompleted too much.

A thread_local version would be another solution, that could be implemented without changing any user source code. Keep the global static interface object for binary compatibility, and move the actual QObject-based implementation to thread_local storage for the gui thread.

The QPixmapCache documentation already says that the cache is only usable from the "application's main thread".

It's an internal class, I didn't intend to modify the QPixmapCache to begin with. (see the woboq link in Chris' post)

I saw, and if I interpreted correctly, the issue is that the global static QPixmapCache and its internal QObject based implementation will be destroyed from the C++ main thread when main() exits. The destruction of the internal implementation object could be removed from the public QPixmapCache destructor, but the question is where to destroy it while maintaining the current intended behavior. Making the internal object (or a wrapper) thread local in the gui thread solves destruction at a point where the cache can't see further use anyway.

[kshegunov]

@jeremy_k said in Global static QPixmapCache in Qt internals:

I saw, and if I interpreted correctly, the issue is that the global static QPixmapCache and its internal QObject based implementation will be destroyed from the C++ main thread when main() exits.

Correct, but only with the allowance that we are talking about the QPMCache instance (which is some internal class).

The destruction of the internal implementation object could be removed from the public QPixmapCache destructor, but the question is where to destroy it while maintaining the current intended behavior.

As I wrote: QCoreApplication::aboutToQuit is one possible such point. The problem with that solution is that it won't set the nullptr of the global pointer, which in turn is going to lead to a crash if the application object is recreated. It shouldn't get recreated, but I've seen code that relies on that (believe or not). With any API the first thing you get is hidden use cases, sadly.

Making the internal object (or a wrapper) thread local in the gui thread solves destruction at a point where the cache can't see further use anyway.

I don't follow. The instance is created dynamically and on demand, and from some thread unknown at the point of declaration. How does specifying a thread local storage helps here, the pointer to the object is global (and I assume used from many a place)?

[jeremy_k]

@kshegunov said in Global static QPixmapCache in Qt internals:

@jeremy_k said in Global static QPixmapCache in Qt internals:

I saw, and if I interpreted correctly, the issue is that the global static QPixmapCache and its internal QObject based implementation will be destroyed from the C++ main thread when main() exits.

Correct, but only with the allowance that we are talking about the QPMCache instance (which is some internal class).

Yes. We seem to be talking around the same point.

The destruction of the internal implementation object could be removed from the public QPixmapCache destructor, but the question is where to destroy it while maintaining the current intended behavior.

As I wrote: QCoreApplication::aboutToQuit is one possible such point. The problem with that solution is that it won't set the nullptr of the global pointer, which in turn is going to lead to a crash if the application object is recreated. It shouldn't get recreated, but I've seen code that relies on that (believe or not). With any API the first thing you get is hidden use cases, sadly.

Making the internal object (or a wrapper) thread local in the gui thread solves destruction at a point where the cache can't see further use anyway.

I don't follow. The instance is created dynamically and on demand, and from some thread unknown at the point of declaration. How does specifying a thread local storage helps here, the pointer to the object is global (and I assume used from many a place)?

Using QCoreApplication::aboutToQuit would change the behavior for applications that repeatedly instantiate and destroy the application instance, and it would add an opportunity to make a mistake for applications that don't use QCoreApplication::exec().

Using a thread local object keeps this into RAII territory. The only difference from the current global static is that destruction happens in the relevant thread rather than in the main() thread.

How does the cache managing object tell that it is in the gui thread? If the first application instance must be created before cache use, QThread::currentThread()->threadId() == QCoreApplication::instance()->thread()->threadId() could work. I don't know if creating the application first is currently required.

[kshegunov]

@jeremy_k said in Global static QPixmapCache in Qt internals:

Using a thread local object keeps this into RAII territory. The only difference from the current global static is that destruction happens in the relevant thread rather than in the main() thread.

Ehm, the object, currently, is created in the heap, is it not? Or perhaps I'm misunderstanding what you mean exactly.

@jeremy_k said in Global static QPixmapCache in Qt internals:

I don't know if creating the application first is currently required.

Yes, it is, but also no QObject shall outlive the QCoreApplication instance is part of the rule, which is exactly what's happening in this case.

PS.

How does the cache managing object tell that it is in the gui thread?

When a QObject is created it stores the reference to the creating thread internally. When the destructor's run that reference's thread id is checked against the current thread id, if they don't match you know you're destroying objects from another thread. This is where this warning is generated.

[jeremy_k]

@kshegunov said in Global static QPixmapCache in Qt internals:

@jeremy_k said in Global static QPixmapCache in Qt internals:

Using a thread local object keeps this into RAII territory. The only difference from the current global static is that destruction happens in the relevant thread rather than in the main() thread.

Ehm, the object, currently, is created in the heap, is it not? Or perhaps I'm misunderstanding what you mean exactly.

IIRC, it is created and destroyed by a QGlobalStatic.

@jeremy_k said in Global static QPixmapCache in Qt internals:

I don't know if creating the application first is currently required.

Yes, it is, but also no QObject shall outlive the QCoreApplication instance is part of the rule, which is exactly what's happening in this case.

The cache object also outlives a stack allocated QCoreApplication when the gui thread is the main thread.

I'm looking for and not finding a reference to the rules about QCoreApplication versus QObject lifetimes. There used to be a concise list, if I didn't imagine it.

[kshegunov]

@jeremy_k said in Global static QPixmapCache in Qt internals:

IIRC, it is created and destroyed by a QGlobalStatic.

Yes, correct. So you want to make the QGlobalStatic instance thread local then?

@jeremy_k said in Global static QPixmapCache in Qt internals:

The cache object also outlives a stack allocated QCoreApplication when the gui thread is the main thread.

Yes, but since they're created and destroyed in the same thread the message doesn't fire up. It's wrong, it's just hidden.

@jeremy_k said in Global static QPixmapCache in Qt internals:

I'm looking for and not finding a reference to the rules about QCoreApplication versus QObject lifetimes. There used to be a concise list, if I didn't imagine it.

I don't remember such a list, but this is a direct quote of Thiago[1]:

And QCoreApplication should be the first QObject created and last destroyed.