Crash using QPainter from secondary thread

jim_kaiser

Will try to repro..

[Edit: Ubuntu 10.04 Qt SDK 1.1.1 Qt 4.7.4.. works.. no crash.. ]

bourbon

Ah, thanks. That's interesting.

giesbert

Sorry, one mistake, I used 4.7.2.
Mingw --> crash, msvs2008 --> works
4.7.0 msvs2008 --> works

very strange...

bourbon

Does this sound like something that should be reported on the Qt bugtracker? If so, where do I do that?

steno

"For info on how to post a bug....":http://developer.qt.nokia.com/wiki/ReportingBugsInQt

andre

[quote author="bourbon" date="1308069402"]Does this sound like something that should be reported on the Qt bugtracker? If so, where do I do that?[/quote]

It does to me. Sounds like a regression, and those are generally taken quite seriously. Don't forget to attach your example to reproduce the issue, link to this topic, and add a tag to this topic with the bug id (QTBUG-#####)

dangelog

The first thing that comes in my mind is that you're not end()ing the painting before releasing the lock, thus you're relying on the QPainter dtor (which will end the painting itself).

There's an obvious race condition there, and I'm confident that helgrind would have told you so. Try adding end() there, or switch to QMutexLocker.

Franzk

The latter will probably give cleaner code.

bourbon

The crash occurs in the very first call to painter.drawLine(), so I don't think it's a race condition.
I didn't realise I shold be calling end(), I've added it but it makes no difference (the code never reaches it).

mlangezaal

I have posted the root cause, a small project to reproduce it and a temporary workaround in the Bugtracker for this bug: QTBUG-19886

It is caused by a stack mis-alignment of the newly created thread. It affects the drawing engine which expects 16-byte alignment for the used SSE2 instructions.
It appears to only occur on windows XP. And happens when a QPainter::drawImage is done on a different thread. The stack on the new thread (QThread subclass and started with start()) is NOT 16-bit aligned, and causes a Segmentation Fault on
the MMX instruction PANDN in qdrawhelper_sse2.cpp.

The only (dirty) workaround I can think of is to manually push extra bytes on the stack in case it is not aligned to realign the SP to a 16-byte boundary. And this seems to work quite well, no more crashes!