<AWT Dev> [8] Review Request: 7090424 TestGlyphVectorLayout failed automately with java.lang.StackOverflowError

Tue Oct 22 08:16:08 PDT 2013

Hi, Anthony.
On 22.10.2013 17:38, Anthony Petrov wrote:
> Hi Sergey,
>
> I'm comparing the changes to XCanvasPeer and XComponentPeer, and I see 
> that previously we would trigger target.repaint() if the current 
> background of the peer is null (regardless of the argument value 
> passed to XCanvasPeer.setBackground().
> The XComponentPeer.setBackground() won'd trigger a repaint if both the 
> current background color and the argument are nulls. Also, it doesn't 
> call target.repaint() directly as the former method in XCanvasPeer 
> did. I see that in XWindow.repaint() there's some logic that might 
> send a repaint event asynchronously depending on the current thread, 
> but still may not do so.
  - This method was added to the XCanvasPeer in this 4947530 to stop 
kind of recursion postRepaint->paint->postRepaint.
  - Usage of target repaint is incorrect here, it should be used from 
the user's code only, because it trigger UPDATE event instead of PAINT 
event.
> Also, XListPeer now dispatches target's paint() asynchronously through 
> an event whereas previously this method has always been called 
> directly. I believe I understand the reason for this: we want to post 
> an event instead of calling the method directly to avoid a 
> StackOverflowError. And this looks fine then.However, in case of the 
> XCanvasPeer I'm not sure we trigger target's repainting in all cases 
> depending on the current thread. I feel there's some change in logic 
> which is not properly documented or may even be a bug that could cause 
> some regressions in user code dealing with canvases.
What does it mean "not sure we trigger target's repainting in all cases 
depending on the current thread"? If background is changed on the EDT we 
paint a native/target, if background changed on other thread we paint 
native part and post PAINT event, which repaint the whole component later.
> Another point is regarding the XWindow.reapint() changes alone. 
> Previously it has called paint(), and the latter simply called 
> paintPeer w/o repainting the target. Only the XComponentPeer repainted 
> the target in its overridden repaint() method. And this does seem 
> logical since it's the XComponentPeer that should maintain the link 
> between a peer and its target. So my question is, would it be more 
> logical to implement the if (dispatchThread) logic in 
> XComponentPeer.paint() instead of XWindow.repaint()?
It is not possible, because a paint() method can be called from the 
Component.paintAll() on any thread and it should paint all parts of the 
component(native/user's) synchronously.
>
> -- 
> best regards,
> Anthony
>
> On 10/22/2013 03:19 PM, Sergey Bylokhov wrote:
>> New version here:
>> http://cr.openjdk.java.net/~serb/7090424/webrev.02
>>
>> On 22.10.2013 13:52, Anthony Petrov wrote:
>>> Hi Sergey,
>>>
>>> Sorry, but I can't review a fix w/o a webrev.
>>>
>>> What kind of problems do you experience with the tool? The cr.openjdk
>>> should now have enough free space to host new webrevs. Please try to
>>> re-upload the new webrev.
>>>
>>> -- 
>>> best regards,
>>> Anthony
>>>
>>> On 10/22/2013 12:28 PM, Sergey Bylokhov wrote:
>>>> I have a problem with the review tool. I assume that version .00 with
>>>> removed volatiles and renamed handleexpose event is ok for everybody?
>>>>
>>>> On 10/21/13 4:53 PM, Sergey Bylokhov wrote:
>>>>> Hi, Anthony.
>>>>> Since cr have some issues with the space. I upload 2 files, where
>>>>> handleExposeEvent was renamed to postPaintEvent():
>>>>>
>>>>> http://cr.openjdk.java.net/~serb/7090424/webrev%2c01/src/solaris/classes/sun/awt/X11/XWindow.java.sdiff.html 
>>>>>
>>>>>
>>>>>
>>>>> http://cr.openjdk.java.net/~serb/7090424/webrev%2c01/src/solaris/classes/sun/awt/X11/XContentWindow.java.sdiff.html 
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On 18.10.2013 15:47, Sergey Bylokhov wrote:
>>>>>> Hi, Anthony.
>>>>>> On 18.10.2013 15:17, Anthony Petrov wrote:
>>>>>>> Hi Sergey,
>>>>>>>
>>>>>>> In XAWT, we usually use the StateLock to synchronize access to peer
>>>>>>> fields (such as background, label, etc.) I don't think that
>>>>>>> switching to volatile is a good idea since it prevents us from
>>>>>>> performing atomic read/writes to the fields. And this is exactly
>>>>>>> what we need for this fix, actually. In other words, the following
>>>>>>> pattern works perfectly:
>>>>>>>
>>>>>>> synchronized (lock) {
>>>>>>>    if (a != b) {
>>>>>>>       a = b;
>>>>>>>       // do stuff, or set a flag to do it later w/o the lock
>>>>>>>    }
>>>>>>> }
>>>>>>>
>>>>>>> whereas the following doesn't:
>>>>>>>
>>>>>>> volatile a;
>>>>>>> if (a != b) {
>>>>>>>    a = b;
>>>>>>>    // do stuff
>>>>>>> }
>>>>>>>
>>>>>>> The latter doesn't work because the value of 'a' may change from
>>>>>>> another thread after the if() statement in the first thread is
>>>>>>> executed.
>>>>>> If it will be changed that's ok. It is safe in this context since
>>>>>> there is no races. a will be the latest setted value and repaint
>>>>>> action will be done after a was set. Non trivial setters(like
>>>>>> setLabel/setText) are called under the locks in shared code.
>>>>>>>
>>>>>>> Please note that this is critical for AWT because it is a
>>>>>>> multi-threaded GUI toolkit.
>>>>>>>
>>>>>>> src/solaris/classes/sun/awt/X11/XListPeer.java
>>>>>>>> - target.paint(g);
>>>>>>>> +                handleExposeEvent(target, 0, 0, getWidth(),
>>>>>>>> getHeight());
>>>>>>>
>>>>>>> (the same applies to XWindow.repaint): can you please rename
>>>>>>> XWindow.handleExposeEvent(Component...) to postPaintEvent() and 
>>>>>>> make
>>>>>>> it final? A good javadoc for this method would also be appreciated,
>>>>>>> because currently seeing the name handle*() I'd think it needs 
>>>>>>> to be
>>>>>>> done under the awtLock().
>>>>>> Will do.
>>>>>>>
>>>>>>> Also, for the corresponding changes in XWindow.repaint(), could you
>>>>>>> please elaborate a bit more? Looking at the code I see that
>>>>>>> XWindow.paint() calls paintPeer(). And in repaint(), you either 
>>>>>>> call
>>>>>>> paint() or paintPeer() depending on the current thread. Why is it
>>>>>>> needed? Can we just call paintPeer() (or paint() for that matter)
>>>>>>> unconditionally since they both seem to result in the same call?
>>>>>> paintPeer is used to draw the native part of the peer(text of the
>>>>>> label, border of the button etc).
>>>>>> paint is a part of Component.paintAll(). And as a result of it call
>>>>>> it should paint the native part of the peer and it should call
>>>>>> appropriate paint method from the target.
>>>>>>> Also, why don't we post an event if reapint() is invoked on the 
>>>>>>> EDT?
>>>>>> We do this in osx, but in xawt there are "smart" caches, which are
>>>>>> initialized during the paint of the peer. See
>>>>>> XLabelPeer&cachedFontMetrics for example . Note that this cache is
>>>>>> completely broken.
>>>>>>>
>>>>>>> -- 
>>>>>>> best regards,
>>>>>>> Anthony
>>>>>>>
>>>>>>> On 10/16/2013 06:00 PM, Sergey Bylokhov wrote:
>>>>>>>> Hello.
>>>>>>>> Please review the fix for jdk 8.
>>>>>>>> The fix has two parts
>>>>>>>>   - Repaint method in the peer now paint the component in place
>>>>>>>> if it
>>>>>>>> was called on EDT only.
>>>>>>>>   - Most of setters were changed to stop recursion if they were
>>>>>>>> called
>>>>>>>> on EDT.
>>>>>>>>
>>>>>>>> Bug: https://bugs.openjdk.java.net/browse/JDK-7090424
>>>>>>>> Webrev can be found at:
>>>>>>>> http://cr.openjdk.java.net/~serb/7090424/webrev.00
>>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>
>>


-- 
Best regards, Sergey.