Hi Anton et al.,

If I read the description of the bug correctly, specifically this part:

The problem occurs, if another thread (for example, GTK thread) is doing the same sort of thing concurrently. This can lead to the following situation. JVM thread: Sets J2DXErrHandler(), saves ANY_PREVIOUS_HANDLER as previous GTK thread: Sets some GTK_HANDLER, saves J2DXErrHandler() as previous JVM thread: Restores ANY_PREVIOUS_HANDLER GTK thread: Restores J2DXErrHandler() JVM thread: Sets J2DXErrHandler(), saves J2DXErrHandler() as previous

It is obvious that at this final step 2D is in an inconsistent state. We don't expect to replace our own error handler (and it shouldn't have been there in the first place).

I realize that the fix you propose works around this problem. But this doesn't look like an ideal solution to me.

BTW, IIRC, in JDK7 (and 6?) we decided to set the actual X11 error handler only once and never replace it. All the rest of the push_handler/pop_handler logic is now located in Java code (see XToolkit.SAVED_ERROR_HANDLER() and the surrounding logic). I believe that we should somehow share this machinery with the 2D code to avoid this sort of problems. Though I'm not sure if this will eliminate this exact issue.


2D/AWT folks: any other thoughts?

--
best regards,
Anthony

On 12/29/2012 17:44, Anton Litvinov wrote:
Hello,

Please review the following fix for a bug.

Bug: http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=8005607
         https://jbs.oracle.com/bugs/browse/JDK-8005607
Webrev: http://cr.openjdk.java.net/~alitvinov/8005607/webrev.00

The bug consists in a crash which is caused by a stack overflow for the reason of an infinite recursion in AWT native function J2DXErrHandler() under certain circumstances on 32-bit Linux OS. The fix is based on introduction of the logic, which detects indirect recursive calls to J2DXErrHandler() by means of a simple counter, to J2DXErrHandler() native function. Such a solution requires minimum code changes, does not alter the handler's code significantly and eliminates this bug.

Adding 2d-dev@openjdk.java.net e-mail alias to the list of recipients of this letter, because the edited function's name is related to Java 2D area of JDK, despite of the fact that the edited file is located in AWT directory.

Thank you,
Anton

Reply via email to