On 29.04.16 18:01, Anton Tarasov wrote:
[CC’ing to 2d-dev to discuss the issue]
On 29 Apr 2016, at 16:14, Sergey Bylokhov <sergey.bylok...@oracle.com> wrote:
On 29.04.16 15:53, Anton Tarasov wrote:
It seems so. But that might be not that critical, because it doesn’t hold (it
won’t) any UI controls and all the UI tree. Anyway it leaks, yes.
Looks like this is crossplatform bug and it also affects d3d/ogl. So probably
we can fix it on the upper level? validatedSrc/Dst/Data is stored the surfaces
which were validated and ready to paint. from the first point of view we can
change them to soft reference, and take care about null values.
Are the validatedSrc/Dst/Data objects referenced from somewhere else? They are
private, so from native? If not, soft refs won’t help I’m afraid...
I guess it is used in BufferedContext only, to skip updates of the
native ogl/d3d context if the target/source surfaces were not changed
since the last update.
This is not the case for CGLLayer, which is referenced from JNI. And so,
wrapping it with a weak ref will work. Also, if the SurfaceData is uniquely
tight to the layer, then it seems natural to dispose (not flush) it with the
layer disposal. And that’s actually the case: LWWindowPeer.disposeImpl()
invalidates it.
But the problem is that invalidation doesn’t release the layer.
Yes, that's right the surface and layer are bound to each other(and the
layer can have more than one surface). So I do not see a reason why we
should break the link between them, which causes the surface to live
more time than its layer. I guess the right things to do is to fix the
"gc root", since we have no cycles here.
So, again the question is: should the layer be nullified on invalidation or it
should be made a weak ref? For me this seems quite logical to release resources
on disposal/invalidation, what do you think?
As to the fixing the issue globally, I don’t have enough understanding of the
pipe design so that to do that properly. For instance, as I wrote before, I
don’t know under which conditions the context should/may be disposed…
May be someone else can advice on it.
I can take a look, but are you sure that the test "WindowsLeak.java"
reproduce exactly your problem?
Note that such changes are 2d related code and should be reviewed on 2d-dev.
Assigned the peer/surfaceData to null in CGLayer can causes an NPE in all its
usage, because there is no any synchronization which will prevent the usage of
CGLayer after disposing.
That’s bad. Will wrapping the refs with AtomicReference help?
Unrelated to the fix, but it seems we should call flush() on surface when the
layer is disposed, at least I do not understand where we flush the native ogl
data for the latest surface data.
This will trigger CGLLayerSurfaceData.invalidate(), but the “layer” will still
not be nullified. What about nullifying it in invalidate()? Will we face the
same synchronisation issue?
Anton.
On 29.04.16 15:00, Anton Tarasov wrote:
Hi Sergey, Alexander,
Please review the fix:
bug: JDK-8028486 [TEST_BUG] [macosx]
java/awt/Window/WindowsLeak/WindowsLeak.java fails
webrev: http://cr.openjdk.java.net/~ant/JDK-8028486/webrev.0
I’m copying my comment from CR:
Please open the attached screenshot [*], made with YourKit, where a chain of
links is shown from the GC roots.
The frame is held by its peer which is held by CGLLayer which is held as
validatedSrcData in the GL context.
The point is that the GL context doesn't cleanup the last state until under
some conditions, which are not applicable to this scenario.
I'm not sure should the cleanup be triggered here or not, but the problem can
be solved otherwise.
The point is that in the chain the CGLLayer instance has been disposed, in
response to the frame disposal.
So, this is the only ref that holds it (the JNI ref is released by the native
peer on disposal).
Thus, as the layer is disposed it can at least zero all the java refs it holds
(this change already fixes the problem).
Then, the "layer" ref in CGLLayerSurfaceData should probably be made weak.
[*] https://bugs.openjdk.java.net/secure/attachment/59121/8028486.png
As to the “weak ref” mentioned in the comment. I didn’t do that, but if you
find it reasonable I can add that change (or file a separate CR).
Also, the fix contains some additional cleanup (not related to this CR): two
more JNI local refs leak, fixed.
Thanks,
Anton.
--
Best regards, Sergey.
--
Best regards, Sergey.
--
Best regards, Sergey.