Segfault when calling QWidget::show (on Debian 9)

J.Hilk

@Bart_Vandewoestyne new debian, huh. New/updated compiler then as well?

I assume you have tried the release build with -O0 ?

Christian Ehrlicher

@J-Hilk or run it with valgrind (compile with -O2 and -g)

J.Hilk

@Christian-Ehrlicher never used valgrind before, as I usually don't do linux stuff. But I trust your expertise :D

oh it also now supports macOS, maybe I should give it a try sometime soon than!

Bart_Vandewoestyne

@J-Hilk said in Segfault when calling QWidget::show (on Debian 9):

@Bart_Vandewoestyne new debian, huh. New/updated compiler then as well?

Yes, due to the switch from Debian 8 to Debian 9, a new compiler as well. Debian 8 (where everything works) has

dev@debian8:~$ g++ --version | head -1
g++ (Debian 4.9.2-10+deb8u2) 4.9.2

while Debian 9 (where the release build segfaults) has

user@debianvbox:~$ g++ --version | head -1
g++ (Debian 6.3.0-18+deb9u1) 6.3.0 20170516

I assume you have tried the release build with -O0 ?

I hadn't, but now I have ;-) And I have interesting news: when using -O0 the segfault is gone! From -O1 and further, we get the segfault.

Bart_Vandewoestyne

@Christian-Ehrlicher said in Segfault when calling QWidget::show (on Debian 9):

@J-Hilk or run it with valgrind (compile with -O2 and -g)

I have no experience with valgrind, but looks like a good suggestion so I will try and report back.

Bart_Vandewoestyne

@Christian-Ehrlicher said in Segfault when calling QWidget::show (on Debian 9):

@J-Hilk or run it with valgrind (compile with -O2 and -g)

OK, so I compiled with -O2 and -g and ran my program through valgrind. This is what I got:

user@debianvbox:~/SVN/PolarisRel/Apps$ valgrind ./PolarisSlave
==5165== Memcheck, a memory error detector
==5165== Copyright (C) 2002-2015, and GNU GPL'd, by Julian Seward et al.
==5165== Using Valgrind-3.12.0.SVN and LibVEX; rerun with -h for copyright info
==5165== Command: ./PolarisSlave
==5165== 
==5165== Invalid read of size 8
==5165==    at 0x8B3B205: XSetCommand (in /usr/lib/x86_64-linux-gnu/libX11.so.6.3.0)
==5165==    by 0x8B3F7EF: XSetWMProperties (in /usr/lib/x86_64-linux-gnu/libX11.so.6.3.0)
==5165==    by 0x5EE707C: QWidgetPrivate::create_sys(unsigned long, bool, bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x5E9F768: QWidget::create(unsigned long, bool, bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x5EA7696: QWidget::setVisible(bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x4FE1FC: show (qwidget.h:497)
==5165==    by 0x4FE1FC: BSPPolarisSlave::mfRun(int, char**, QString&) (BSPPolarisSlave.cpp:443)
==5165==    by 0x7B50D9: ICService::mfExec(int, char**, QString&, bool) (in /home/user/SVN/PolarisRel/Apps/PolarisSlave)
==5165==    by 0x7B190D: ICService::mfParseArguments(int, char**, bool) (in /home/user/SVN/PolarisRel/Apps/PolarisSlave)
==5165==    by 0x501BB2: BSPPolarisSlave::mfParseArguments(int, char**) (BSPPolarisSlave.cpp:659)
==5165==    by 0x4DE3AC: main (BSPPolarisSlaveMain.cpp:71)
==5165==  Address 0xbe71780 is 0 bytes after a block of size 16 alloc'd
==5165==    at 0x4C2C93F: operator new[](unsigned long) (vg_replace_malloc.c:423)
==5165==    by 0x501A27: BSPPolarisSlave::mfParseArguments(int, char**) (BSPPolarisSlave.cpp:637)
==5165==    by 0x4DE3AC: main (BSPPolarisSlaveMain.cpp:71)
==5165== 
==5165== Invalid read of size 1
==5165==    at 0x4C2EDA2: strlen (vg_replace_strmem.c:454)
==5165==    by 0x8B3B1EC: XSetCommand (in /usr/lib/x86_64-linux-gnu/libX11.so.6.3.0)
==5165==    by 0x8B3F7EF: XSetWMProperties (in /usr/lib/x86_64-linux-gnu/libX11.so.6.3.0)
==5165==    by 0x5EE707C: QWidgetPrivate::create_sys(unsigned long, bool, bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x5E9F768: QWidget::create(unsigned long, bool, bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x5EA7696: QWidget::setVisible(bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x4FE1FC: show (qwidget.h:497)
==5165==    by 0x4FE1FC: BSPPolarisSlave::mfRun(int, char**, QString&) (BSPPolarisSlave.cpp:443)
==5165==    by 0x7B50D9: ICService::mfExec(int, char**, QString&, bool) (in /home/user/SVN/PolarisRel/Apps/PolarisSlave)
==5165==    by 0x7B190D: ICService::mfParseArguments(int, char**, bool) (in /home/user/SVN/PolarisRel/Apps/PolarisSlave)
==5165==    by 0x501BB2: BSPPolarisSlave::mfParseArguments(int, char**) (BSPPolarisSlave.cpp:659)
==5165==    by 0x4DE3AC: main (BSPPolarisSlaveMain.cpp:71)
==5165==  Address 0x50 is not stack'd, malloc'd or (recently) free'd
==5165== 
==5165== 
==5165== Process terminating with default action of signal 11 (SIGSEGV)
==5165==  Access not within mapped region at address 0x50
==5165==    at 0x4C2EDA2: strlen (vg_replace_strmem.c:454)
==5165==    by 0x8B3B1EC: XSetCommand (in /usr/lib/x86_64-linux-gnu/libX11.so.6.3.0)
==5165==    by 0x8B3F7EF: XSetWMProperties (in /usr/lib/x86_64-linux-gnu/libX11.so.6.3.0)
==5165==    by 0x5EE707C: QWidgetPrivate::create_sys(unsigned long, bool, bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x5E9F768: QWidget::create(unsigned long, bool, bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x5EA7696: QWidget::setVisible(bool) (in /home/user/SVN/PolarisRel/ThirdParty/Qt/qt-install/lib/libQtGui.so.4.8.7)
==5165==    by 0x4FE1FC: show (qwidget.h:497)
==5165==    by 0x4FE1FC: BSPPolarisSlave::mfRun(int, char**, QString&) (BSPPolarisSlave.cpp:443)
==5165==    by 0x7B50D9: ICService::mfExec(int, char**, QString&, bool) (in /home/user/SVN/PolarisRel/Apps/PolarisSlave)
==5165==    by 0x7B190D: ICService::mfParseArguments(int, char**, bool) (in /home/user/SVN/PolarisRel/Apps/PolarisSlave)
==5165==    by 0x501BB2: BSPPolarisSlave::mfParseArguments(int, char**) (BSPPolarisSlave.cpp:659)
==5165==    by 0x4DE3AC: main (BSPPolarisSlaveMain.cpp:71)
==5165==  If you believe this happened as a result of a stack
==5165==  overflow in your program's main thread (unlikely but
==5165==  possible), you can try to increase the size of the
==5165==  main thread stack using the --main-stacksize= flag.
==5165==  The main thread stack size used in this run was 8388608.
==5165== 
==5165== HEAP SUMMARY:
==5165==     in use at exit: 1,121,308 bytes in 9,104 blocks
==5165==   total heap usage: 22,310 allocs, 13,206 frees, 4,039,124 bytes allocated
==5165== 
==5165== LEAK SUMMARY:
==5165==    definitely lost: 2,944 bytes in 6 blocks
==5165==    indirectly lost: 13,190 bytes in 537 blocks
==5165==      possibly lost: 54,718 bytes in 437 blocks
==5165==    still reachable: 1,050,456 bytes in 8,124 blocks
==5165==         suppressed: 0 bytes in 0 blocks
==5165== Rerun with --leak-check=full to see details of leaked memory
==5165== 
==5165== For counts of detected and suppressed errors, rerun with: -v
==5165== ERROR SUMMARY: 5 errors from 2 contexts (suppressed: 0 from 0)
Segmentation fault

I'll try to decipher this myself, but if in the meanwhile someone more experienced with valgrind can point me in the right direction, that would be nice :-)

J.Hilk

from my point of view, I would say the issue is with your strange string manipulation stuff.

You have to be very careful, when working with stringliterals, it is super easy to run into undefined behaviour, when you try to modify them.

Christian Ehrlicher

@Bart_Vandewoestyne said in Segfault when calling QWidget::show (on Debian 9):

by 0x501A27: BSPPolarisSlave::mfParseArguments(int, char**) (BSPPolarisSlave.cpp:637)

This is where you have to take a look on. You do something wrong there for an argument.

kshegunov

Please provide the the line that @Christian-Ehrlicher mentioned; it's the call to mfExec.

Bart_Vandewoestyne

OK, I think we're getting there... In ICBlackBoxBase::mfInitialize we call the ICBlackBoxBaseApplication constructor which calls the QApplication constructor with certain argc and argv arguments:

ICBlackBoxBaseApplication::ICBlackBoxBaseApplication(int &argc, char** argv, ICBlackBoxBase* apApp)
: QApplication(argc,argv), mpApp(apApp)
{
}

Now let's see what argc and argv we are passing there. I've set a breakpoint right before the location where we call this constructor, and this is the call stack:

(gdb) bt
#0  ICBlackBoxBase::mfInitialize (this=0x7fffffffe0e0, argc=2, argv=0x5555565cd3d0, errormsg=...) at ICBlackBoxBase.cpp:101
#1  0x0000555555c0101d in ICService::mfExec(int, char**, QString&, bool) ()
#2  0x0000555555bfd86e in ICService::mfParseArguments(int, char**, bool) ()
#3  0x000055555594db13 in BSPPolarisSlave::mfParseArguments (this=0x7fffffffe0e0, argc=2, argv=0x7fffffffe258) at BSPPolarisSlave.cpp:654
#4  0x000055555592a3bd in main (argc=1, argv=0x7fffffffe258) at BSPPolarisSlaveMain.cpp:71

As you can see, in main we have that argc is 1, but in ICBlackBoxBase::mfInitialize (the function from which we call the ICBlackBoxBaseApplication constructor, and thus also the QApplication constructor) we have that argc is 2 (since an extra -e argument was added). Now let's look at argv in both main and ICBlackBoxBase::mfInitialize. In main we have:

(gdb) f 4
#4  0x000055555592a3bd in main (argc=1, argv=0x7fffffffe258) at BSPPolarisSlaveMain.cpp:71
71		return (polarisSlave.mfParseArguments(argc, argv));	
(gdb) p argc
$12 = 1
(gdb) p argv[0]
$13 = 0x7fffffffe53a "/home/user/SVN/PolarisRel/Apps/PolarisSlave"
(gdb) p argv[argc]
$14 = 0x0

but in ICBlackBoxBase::mfInitialize we have:

(gdb) f 0
#0  ICBlackBoxBase::mfInitialize (this=0x7fffffffe0e0, argc=2, argv=0x5555565cd3d0, errormsg=...) at ICBlackBoxBase.cpp:101
101	{
(gdb) p argc
$15 = 2
(gdb) p argv[0]
$16 = 0x5555565cf830 "/home/user/SVN/PolarisRel/Apps/PolarisSlave"
(gdb) p argv[1]
$17 = 0x5555565cf940 "-e"
(gdb) p argv[argc]
$18 = 0x20 <error: Cannot access memory at address 0x20>

so there argv[argc] is not null! And now I have to find out why :-)

JonB

@Bart_Vandewoestyne said in Segfault when calling QWidget::show (on Debian 9):

so there argv[argc] is not null! And now I have to find out why :-)

Earlier I wrote:

I shall be surprised if it is this, but....

I think your code is not 100% technically correct. You do not NULL terminate your new vector. Technically you should find your original argv had an extra element at the end: argv[argc] == NULL. You do not copy this or NULL terminate your new newArgvs. E.g. https://stackoverflow.com/questions/16418932/is-argvargc-equal-to-null-pointer
— argv[argc] shall be a null pointer.

It is not clear whether this matters or not. If code only uses argc to index up to argv[argc - 1] then it does not. If code does do something about looking at argv[argc] to check for nullptr then it does matter. If you have the source code where it goes wrong you may be able to delermine.

JonB

@Bart_Vandewoestyne said in Segfault when calling QWidget::show (on Debian 9):

ICBlackBoxBaseApplication::ICBlackBoxBaseApplication(int &argc, char** argv, ICBlackBoxBase* apApp)
: QApplication(argc,argv), mpApp(apApp)
{
}

I have an alternative theory. Work with me on this one, please!

I believe the BSPPolarisSlave & ICService classes are your own(?). This ICBlackBoxBaseApplication constructor calls QApplication(argc,argv) with argc, and especially argv, *with the values they have on entry to the constructor. However (I believe you are saying) something it then does alters argc, and especially argv, to add further arguments, right? Note in your traceback how argv has a different (pointer) value from BSPPolarisSlave::mfParseArguments() down to ICBlackBoxBase::mfInitialize() --- it has been changed.

Now, I am unclear whether the original argv, with its strings, is being used from the original QApplication(argc,argv), and just possibly the value/strings array is no longer valid. But it may be what is used the very first time QWidget::show() is called (to do some X initialisations from any (potential) X-type command-line arguments passed to your Qt app. Remember that https://doc.qt.io/qt-5/qcoreapplication.html#QCoreApplication says

Warning: The data referred to by argc and argv must stay valid for the entire lifetime of the QCoreApplication object. In addition, argc must be greater than zero and argv must contain at least one valid character string.

I think you should find that QWidget::show() which is in BSPPolarisSlave::mfRun(). On the line above it do something about printing out Qt's idea of the command line arguments which were used during the earlier QApplication(argc,argv). I believe static QStringList QCoreApplication::arguments() accesses these. So qDebug() those before QWidget::show(), do they look good or garbled?

Finally, the question is why/what are these classes altering argc/argvfor, apparently to add additional arguments? If these new arguments are intended to be seen by Qt then it looks like you should not be callingQApplication(argc,argv)` until after they have been modified to hold the new command line?

Bart_Vandewoestyne

I have also found the following text on https://doc.qt.io/archives/qt-4.8/qapplication.html#details

"Since the QApplication object does so much initialization, it must be created before any other objects related to the user interface are created. QApplication also deals with common command line arguments. Hence, it is usually a good idea to create it before any interpretation or modification of argv is done in the application itself."

It worries me... I don' think we actually modify argv (but we do modify argc)... but we do create a newArgv that has one more command line argument and that newArgv together with the modified argc is what gets passed to QApplication later on...

So in summary, I think this is what we do:

Start from main with argc and argv.
increment argc by one and create newArgv from argv (newArgv has one extra -e command line argument).
Call the QApplication constructor with the incremented argc and newArgv...

JonB

@Bart_Vandewoestyne
I know this. I suggested what you need to do, to the QWidget::show() in BSPPolarisSlave::mfRun()....

Bart_Vandewoestyne

@JonB said in Segfault when calling QWidget::show (on Debian 9):

@Bart_Vandewoestyne
I know this. I suggested what you need to do, to the QWidget::show() in BSPPolarisSlave::mfRun()....

Yes, thanks for that suggestion. I'll look into it. I guess we were both typing our replies at the same time, but you were the first to press the 'Submit' button :-) Stay tuned for more! :-)

JonB

@Bart_Vandewoestyne
Hold on.

First there is a logic fault in new code. Although it may not matter to your error right now, because you are still inside ICService::mfParseArguments(argc, newArgs.data()), your newArgs (local) variable does not respect

The data referred to by argc and argv must stay valid for the entire lifetime of the QCoreApplication object

At least try making newArgs static.

You still have not said "who" is supposed to "see" these extra arguments --- QCoreApplication or not?

Also, who uses the -e you append, and are you sure what uses that does not expect a further argument to follow the -e? [EDIT Oh, I see you insert the -e, not append it.]

Bart_Vandewoestyne

@JonB said in Segfault when calling QWidget::show (on Debian 9):

I believe the BSPPolarisSlave & ICService classes are your own(?).

For the record: yes. I even think the ICService class is based on an older version of QtService from https://skycoder42.github.io/QtService/.

This ICBlackBoxBaseApplication constructor calls QApplication(argc,argv) with argc, and especially argv, *with the values they have on entry to the constructor. However (I believe you are saying) something it then does alters argc, and especially argv, to add further arguments, right?

I think it's the other way around: we create a newArgv that has one extra -e parameter and we also increment argc, and those are eventually passed to QApplication.

Note in your traceback how argv has a different (pointer) value from BSPPolarisSlave::mfParseArguments() down to ICBlackBoxBase::mfInitialize() --- it has been changed.

OK. This indeed sounds not good... I will look into that.

[...]
I think you should find that QWidget::show() which is in BSPPolarisSlave::mfRun(). On the line above it do something about printing out Qt's idea of the command line arguments which were used during the earlier QApplication(argc,argv). I believe static QStringList QCoreApplication::arguments() accesses these. So qDebug() those before QWidget::show(), do they look good or garbled?

I've changed

mpMainWindow->show();

into

QApplication::arguments();
mpMainWindow->show();

and when I run the program in gdb, it now no longer crashes at the call to show(), but rather at the call to QApplication::arguments(), again pointing in the direction that clearly something is wrong with this argc, argv and newArgvs code...

Finally, the question is why/what are these classes altering argc/argvfor, apparently to add additional arguments? If these new arguments are intended to be seen by Qt then it looks like you should not be callingQApplication(argc,argv)` until after they have been modified to hold the new command line?

This is the code that does the modification:

	//As PolarisSlave is derived from ICService we must start with -e argument to run
	//PolarisSlave as a regular program and not as a service.
	char** newArgvs;
	newArgvs = new char*[argc+1];
	for(i = 0; i < argc + 1; i++)
	{
		newArgvs[i] = new char[256];
	}

	strcpy(newArgvs[0], argv[0]);

	if (!showHelp)
	{
		strcpy(newArgvs[1], "-e");		
		for(i = 1; i < argc; i++)
		{
			strcpy(newArgvs[i+1], argv[i]);
		}
		argc++;
	}
	else
	{
		strcpy(newArgvs[1], "-h");
	}

        return ICService::mfParseArguments(argc, newArgvs, false);

Note that we first do the argc/argv/newArgvs stuff, and then we call the function

return ICService::mfParseArguments(argc, newArgvs, false);

which passes these arguments on to other functions and eventually calls the QApplication constructor with these modified arguments.

Also, reading the first comment, I would say that that -e argument is added because we don't want to start the program as a service (remember our ICService class is probably based on QtService, and using -e seems to avoid that we run it as a service).

JonB

@Bart_Vandewoestyne said in Segfault when calling QWidget::show (on Debian 9):

QApplication::arguments();

I said to try:

qDebug() << QApplication::arguments();

though of course it will crash.

Your code looks dodgy to me, at least in the showHelp case, which probably doesn't arise in your current example. I'm not going to type up the potential issues you might encounter, use a debugger if you want to step through. I don't know how that relates to the issue of the QApplication arguments anyway. I don't know why you insert a -e in this code when you previously showed newArgs.push_back(const_cast<char*>("-e")); elsewhere.

Basically I am finding it too difficult to follow your code, and I'm not sure if you are changing stuff. In any case this should be a debugging exercise now. I can see that the first QWidget call needs QApplication::arguments() to be correct, and I think that is not, so follow how that goes from valid to inavlid through your sequence of calls/changes.

Bart_Vandewoestyne

@JonB said in Segfault when calling QWidget::show (on Debian 9):

I said to try:

qDebug() << QApplication::arguments();

I tried that, but I didn't get any output except for a segmentation fault. And that looked rather reasonable to me because I assume QApplication::arguments() gets called before operator<<, not? And if QApplication::arguments() segfaults, we won't see any further output...

Bart_Vandewoestyne

@JonB said in Segfault when calling QWidget::show (on Debian 9):

[...]
Basically I am finding it too difficult to follow your code, and I'm not sure if you are changing stuff. In any case this should be a debugging exercise now. I can see that the first QWidget call needs QApplication::arguments() to be correct, and I think that is not, so follow how that goes from valid to inavlid through your sequence of calls/changes.

No problem. Thanks to you (and the others in this thread) for your help! I now have a clearer idea of where the problem is. It looks like it has indeed something to do with that modification of argc and argv. I'm used to debugging with the Visual Studio debugger, but this one has to be tackled in Linux, so I'll probably have to improve my gdb skills a bit... ;-)