Non-Reentrant Class Use In Multithreaded Program

JKSH

@Crag_Hack said in Non-Reentrant Class Use In Multithreaded Program:

I forgot to clarify in the previous post not just two objects isolated to their own threads but two objects of two different classes... But you knew that right?

Yes, I knew that :)

The widget example still stands: Suppose you create a very simple program that consists of only 1 QApplication object and 1 QWidget object. These two objects are not allowed to be in different threads.

Anything else to know? Are these classes the exception, perhaps only graphical classes?

None that I can think of. I'm quite sure it's only the graphical classes with QApplication/QGuiApplication.

the fileListCompare doesn't use global data (i.e. static/global variables) - it's reentrant - or if it does, then all accesses to the globals is serialized between different threads (with a mutex) - it's thread-safe.

It only uses the referance parameters passed to the function.

And of course this sounds good as well right...

You said: fileListCompare() only uses references parameters. I presume you were confirming that fileListCompare() doesn't use global/static variables.

As @kshegunov said: If fileListCompare() doesn't use global/static variables, then fileListCompare() is a reentrant function.

Crag_Hack

A couple more questions creeped into my domain... thankfully they're quick :) is qDebug() reentrant or should it not be used from multiple threads? All STL classes are reentrant right except for <atomic> and std::atomic guys right?

JKSH

@Crag_Hack said in Non-Reentrant Class Use In Multithreaded Program:

A couple more questions creeped into my domain... thankfully they're quick :) is qDebug() reentrant or should it not be used from multiple threads?

qDebug() itself is thread-safe. It can be called simultaneously from multiple threads, without endangering your application.

However, qDebug() sends data to stderr by default, which is unbuffered. That means: If you send a long stream of "A"s from one thread and a long stream of "B"s from another thread, you might see interleaved text when viewing stderr:

"AAAAAAAAAABBBBBBBBBBAAAAAAAAAABBBBBBBBBB"

To avoid this, call qInstallMessageHandler() to send qDebug() output to a buffered stream (like stdout or a file).

All STL classes are reentrant right except for <atomic> and std::atomic guys right?

I'd say so

Crag_Hack

Another related question I conjured up - how come some Qt classes aren't reeentrant that would be quite more useful if they were? I.E. QPair, QStorageInfo. My program makes use of drive storage info in both the main GUI thread and the worker thread (I was forced to use the win32 api in the worker thread since QStorageInfo wasn't reentrant) and I could easily foresee QPair being used in such a way as well.
Thanks ! That might finalize this thread permanently :)

kshegunov

QPair is reentrant, if the docs don't state it, then it's a "bug" in the documentation, please file it as such.

As for QStorageInfo - I don't know if it's reentrant or not, I would have to check the source to find out for sure, but usually the class not being reentrant is due to the underlying API architecture (e.g. the widget classes).

Crag_Hack

Thanks. Perhaps they are both bugs - I would think a useful class that would have any possibility of being used in multiple threads like QStorageInfo would most definitely be reentrant. Is there another factor at play here besides utility?

kshegunov

@Crag_Hack said in Non-Reentrant Class Use In Multithreaded Program:

utility

What do you mean by "utility"? Whether it's reentrant or thread-safe (or both or neither) just depends on the specific implementation. As Qt's (mostly) a platform agnostic toolkit, implementations for some platforms may be reentrant/thread-safe, while implementations for other platforms may be not, in that case the worst-case is documented - i.e. non reentrant/not thread-safe.

kshegunov

After a quick (not exhaustive) look here:
https://code.woboq.org/qt5/qtbase/src/corelib/io/qstorageinfo.h.html
https://code.woboq.org/qt5/qtbase/src/corelib/io/qstorageinfo.cpp.html
https://code.woboq.org/qt5/qtbase/src/corelib/io/qstorageinfo_p.h.html
https://code.woboq.org/qt5/qtbase/src/corelib/io/qstorageinfo_unix.cpp.html
https://code.woboq.org/qt5/qtbase/src/corelib/io/qstorageinfo_win.cpp.html

It seems to me QStorageInfo is reentrant on *nix, but is not on windows (due to ::SetErrorMode). I didn't look at macos' implementation.

Crag_Hack

Why doesn't the Qt Company make a class like QStorageInfo reentrant though? It would seem logical to do so since it's an easy situation to find yourself in needing to use the class in multiple threads.

JKSH

@Crag_Hack said in Non-Reentrant Class Use In Multithreaded Program:

Why doesn't the Qt Company make a class like QStorageInfo reentrant though? It would seem logical to do so since it's an easy situation to find yourself in needing to use the class in multiple threads.

No strong reason why.

You can post this suggestion to https://bugreports.qt.io/

kshegunov

@JKSH said in Non-Reentrant Class Use In Multithreaded Program:

No strong reason why.

I mentioned already, that ::SetErrorMode breaks reentrancy.

JKSH

@kshegunov said in Non-Reentrant Class Use In Multithreaded Program:

I mentioned already, that ::SetErrorMode breaks reentrancy.

Yes, you're right.

The current implementation is non-reentrant, but I wonder if it's possible to remove this dependency on SetErrorMode(). MSDN recommends SetThreadErrorMode() instead of SetErrorMode() -- I'm not familiar with the Windows API though, so I don't know the implications of this change.

kshegunov

@JKSH said in Non-Reentrant Class Use In Multithreaded Program:

possible to remove this dependency

Maybe, maybe not. I don't know what was the intent of the developer that implemented it. He/She may've had good reason to use this, or may not. In any case this may solve the reentrancy only partially. I haven't looked at the macOS's implementation at all, which may very well require a global state, and to be frank only glanced through the *nix and windows ones.

Perhaps it's better to start with a question to the mailing/development list why it was implemented like this, if someone remembers. But I guess a suggestion wouldn't hurt, so perhaps do them in parallel?

Crag_Hack

About to go live in the next week with one last thing to do... I am going to use the Win API for file copy instead of QFile since I can turn off write buffering. To track progress of the file copy I need to pass a callback to the copy function. See here (the lpProgressRoutine guy). The callback needs to access two worker thread object member functions. My worker thread is simply an object I pass to moveToThread:

    workerThread = new QThread(this);
    workerThread->start();
    worker = new Worker;
    worker->moveToThread(workerThread);

class Worker : public QObject
{
    Q_OBJECT
...
all worker thread code here

What's the best way to do this in a thread-safe manner? Here Kuba simply makes the callback function a static class member fuction.

static DWORD CALLBACK copyProgress(...

Is that safe for me to do as long as the static callback function only accesses the reentrant worker thread object? The current code follows. I need to access the setProgressMeters and setTimeLabels functions from the callback function - I can make jobSize and startTime member variables of the worker thread object and readSize and fileSize equivalent variables will be passed to the callback function as part of the CopyFileEx setup.

void Worker::jobCopyFiles(BackupJob &job, QTime &startTime)
{
...
    QByteArray buffer;
    for (int count = 0; !(buffer = sourceFile.read(1000000)).isEmpty() && cancel == false; count++)
    {
        int readSize = buffer.size();
        targetTempFile.write(buffer);
        setProgressMeters(readSize, fileSize, jobSize);
        setTimeLabels(startTime.elapsed(),readSize);
    }

Please let me know if you need more information. Thanks!

SGaist

What parameters can you pass to that callback ?

Crag_Hack

You can pass via the lpData parameter to the CopyFileEx function. lpDada is LPVOID which is this according to msdn:

A pointer to any type.
This type is declared in WinDef.h as follows:
typedef void *LPVOID;

SGaist

So you can pass a structure or an object that will be available in your callback after casting the pointer to a suitable class.

You can then use QMetaObject::invokeMethod to avoid trouble with respect to the event loop.

Crag_Hack

Do I make the callback function a static worker object class member? Or do I declare it as a global outside of the class? I can't do a normal worker object class member function and still pass it as a callback to the CopyFileEx function right? What's the best way?

Can you show me some code? I'm not terribly well rehearsed in these realms...
Thanks

SGaist

A function outside your class, that callback is an "implementation detail", it should go into your implementation file. Pass your object as parameter to it.
WARNING: Not tested, nor compiled

DWORD CALLBACK CopyProgressRoutine(
	LARGE_INTEGER TotalFileSize,
	LARGE_INTEGER TotalBytesTransferred,
	LARGE_INTEGER StreamSize,
	LARGE_INTEGER StreamBytesTransferred,
	DWORD dwStreamNumber,
	DWORD dwCallbackReason,
	HANDLE hSourceFile,
	HANDLE hDestinationFile,
	LPVOID lpData
	)
{
    int percentage = (double(TotalBytesTransferred.QuadPart) / double(TotalFileSize.QuadPart)) * 100;
    MyCoolClass *object = qobject_cast<MyCoolClass *>(lpData);
    if(object) {
        QMetaObject::invokeMethod(object, 
                                  "nameOfYourMethod", 
                                  Qt::QueuedConnection,
                                  Q_ARG(int, percentage));
    }
    return PROGRESS_CONTINUE;
}


void YouCoolClass::copyFile(const QString& source, const QString& destination)
{
        QByteArray srcUtf8 = QDir::toNativeSeparator(source).toUtf8();
        char *srcFilename = srcUtf8.data();
        QByteArray dstUtf8 = QDir::toNativeSeparator(destination).toUtf8();
        char *dstFilename = dstUtf8.data();
       
	bool returnVal;

	returnVal = CopyFileEx(
		srcFilename,
		dstFilename,
		(LPPROGRESS_ROUTINE)CopyProgressRoutine,
		this,
		NULL,
		COPY_FILE_NO_BUFFERING);

	if(returnVal){
		printf("%s copied to current directory.\n", filename);
	} else {
		printf("%s not copied to current directory.\n", filename);
		printf("Error %u.\n", GetLastError());
	}
}

Crag_Hack

Thanks SGaist! The thing other than static/global/worker member I wasn't sure about was how to get that callback function to call those worker thread object functions. So we have to use QMetaObject::invokeMethod? We can't call the functions directly? What's the reason for that? Is CopyFileEx blocking synchronous? If so wouldn't that prevent queued events from being processed?