Bug 166

Summary: Memory leak with ATI Mobility Radeon 9700
Product: [JogAmp] Jogl Reporter: Sven Gothel <sgothel>
Component: coreAssignee: Sven Gothel <sgothel>
Status: VERIFIED FIXED    
Severity: normal    
Priority: P3    
Version: 1   
Hardware: All   
OS: all   
Type: DEFECT SCM Refs:
Workaround: ---
Attachments: Main program of test case
GLEventListener for test case

Description Sven Gothel 2010-03-24 07:47:54 CET


---- Reported by kbr 2005-06-22 00:25:22 ----

A very simple test case reported by a user on the javagaming.org JOGL forum was
shown to leak memory when run with JOGL 1.1 b12 on ATI Mobility Radeon 9700
hardware on Windows XP with recent drivers. The test case is attached and can
also be found in the following thread on javagaming.org (called "JOGL is slowly
increasing memory usage! WHY?":
http://192.18.37.44/forums/index.php?topic=9865.0

The same test case when ported to LWJGL does not exhibit the memory leak.

The difference in behavior lies in JOGL's and LWJGL's OpenGL context management.
When JOGL performs an OpenGL repaint cycle in the GLCanvas (on Windows), it
locks the surface using the JAWT, makes the OpenGL context current on the
resulting HDC, performs the user's rendering, releases the current OpenGL
context, and unlocks the surface. The current implementation of LWJGL behaves
similarly, but does not release the current OpenGL context before unlocking the
surface, and before making the OpenGL context current on the HDC checks to see
whether it is already current. This skips the "makeCurrent/free" pattern of the
OpenGL context on all but the first paint cycle, possibly for performance
reasons. It can be argued that this OpenGL context management is not completely
correct, because if the OpenGL context was desired to be operated upon from
another thread it would have to cooperatively be released from the thread upon
which it was previously made current. In the general case this is not possible.

It turns out there is a real bug in ATI's drivers causing memory to be leaked on
this particular mobile chip (and perhaps others) during each makeCurrent/free
cycle. This is the root cause of the memory leak with the JOGL program and why
the LWJGL port of the same program does not leak memory.



---- Additional Comments From kbr 2005-06-22 00:26:13 ----

Created an attachment
Main program of test case




---- Additional Comments From kbr 2005-06-22 00:26:59 ----

Created an attachment
GLEventListener for test case




---- Additional Comments From kbr 2005-06-22 01:06:20 ----

Worked around memory leak in ATI's OpenGL drivers by adding system property
-Djogl.GLContext.nofree which users can specify on the command line. There is no
good general-purpose workaround for this bug which works well on all hardware
and in all kinds of applications. Issues may remain if this workaround is used
and if the GLCanvas is removed and re-added to its parent container. Use at your
own risk.




---- Additional Comments From kbr 2005-06-22 01:11:01 ----

I should also mention that a bug has been filed with ATI's Customer Support
Ticket system.




--- Bug imported by sgothel@jausoft.com 2010-03-24 07:47 EDT  ---

This bug was previously known as _bug_ 166 at https://jogl.dev.java.net/bugs/show_bug.cgi?id=166
Imported an attachment (id=59)
Imported an attachment (id=60)

The original submitter of attachment 59 [details] is unknown.
   Reassigning to the person who moved it here: sgothel@jausoft.com.
The original submitter of attachment 60 [details] is unknown.
   Reassigning to the person who moved it here: sgothel@jausoft.com.