Hi Alexandr,
I created a new patch v8 @
http://cr.openjdk.java.net/~dingxmin/7189299/webrev.08/. It uses the
newly proposed approach mentioned in my last email. Could you please
help to review it again?
The patch, of course, passed jtreg test bug7189299.java in the
patch. What's more, I did additional tests for JComboBox, JTextField
and JList in my IDE by comparing listener numbers observed during
debugging with/without my patch. The listener numbers were doubled
without the fix. This proves that v8 patch works for all components. I
think I can write those additional tests as jtreg tests after the patch
passes review.
One notable change is that I had to restrict the scope of holding a
write lock in DefaultStyledDocument because otherwise, we cannot store
the new component created in FormView.createComponent(). The stack
trace is pasted below for reference.
Exception in thread "main" java.lang.reflect.InvocationTargetException
at java.awt.EventQueue.invokeAndWait(EventQueue.java:1270)
at javax.swing.SwingUtilities.invokeAndWait(SwingUtilities.java:1350)
at bug7189299.main(bug7189299.java:116)
Caused by: java.lang.IllegalStateException: Attempt to mutate in
notification
at
javax.swing.text.AbstractDocument.writeLock(AbstractDocument.java:1357)
at
javax.swing.text.html.HTMLDocument.obtainLock(HTMLDocument.java:1708)
at javax.swing.text.html.FormView.createComponent(FormView.java:211)
......(omitted)......
at
javax.swing.plaf.basic.BasicTextUI$RootView.insertUpdate(BasicTextUI.java:1602)
at
javax.swing.plaf.basic.BasicTextUI$UpdateHandler.insertUpdate(BasicTextUI.java:1861)
at
javax.swing.text.AbstractDocument.fireInsertUpdate(AbstractDocument.java:201)
at
javax.swing.text.DefaultStyledDocument.create(DefaultStyledDocument.java:155)
<------ FormView.createComponent is triggered by this call which has
already held the lock
at javax.swing.text.html.HTMLDocument.create(HTMLDocument.java:469)
......(omitted)......
This change did violate what is documented in AbstractDocument.writeLock
that "This situation violates the bean event model where order of
delivery is not guaranteed and all listeners should be notified before
further mutations are allowed. " However, without the change of
shrinking lock scope, the component cannot be saved once it is created
in FormView.createComponent(). I found it is even harder to save it
later in code but perhaps there does exist a more appropriate place to
do this. If you have any better suggestion, I am glad to know.
Also I searched references to 'StyleConstants.ComponentAttribute' as
you asked. The result is listed below. Note that the command 'grep' is
invoked under openjdk 8 top directory.
$ grep -R 'ComponentAttribute' .
./jdk/src/share/classes/javax/swing/text/StyleConstants.java: public
static final Object ComponentAttribute = new
CharacterConstants("component");
./jdk/src/share/classes/javax/swing/text/StyleConstants.java: return
(Component) a.getAttribute(ComponentAttribute);
./jdk/src/share/classes/javax/swing/text/StyleConstants.java:
a.addAttribute(ComponentAttribute, c);
./jdk/src/share/classes/javax/swing/text/StyleConstants.java:
Background, ComponentAttribute, IconAttribute,
./jdk/src/share/classes/javax/swing/text/StyledEditorKit.java:
set.removeAttribute(StyleConstants.ComponentAttribute);
Binary file ./jdk/test/sun/tools/jhat/jmap.bin matches
So ComponentAttribute is not referenced by other classes except
StyledEditorKit. StyleConstants exposes ComponentAttribute in
getComponent() and setComponent() methods. References to
StyleConstants.getComponent and StyleConstants.setComponent are further
searched in repo.
$ grep -R 'StyleConstants.getComponent' .
./jdk/src/share/classes/javax/swing/text/ComponentView.java: * the
element (by calling StyleConstants.getComponent). A
./jdk/src/share/classes/javax/swing/text/ComponentView.java: Component
comp = StyleConstants.getComponent(attr);
$ grep -R 'StyleConstants.setComponent' .
./jdk/src/share/classes/javax/swing/JTextPane.java:
StyleConstants.setComponent(inputAttributes, c);
From the facts above, I think we are sufficiently confident to use
ComponentAttribute.
Please let me know if you have any comments and I can do more regression
tests and provide more jtreg test cases as next step.
Thanks and regards,
Frank
On 11/15/2012 5:18 PM, Alexander Scherbatiy wrote:
On 11/14/2012 10:22 AM, Frank Ding wrote:
Hi Alexandr,
After a couple of hours investigating the possibility of fixing
JComboBox.setModel(null) and JTextComponent.setComponent(null), I
found that
1. In JComboBox.setModel, the new model, null in this case, is
eventually passed to BasicComboPopup.propertyChange where
JList.setModel is called. JList.setModel rejects the null model
because of its api restriction. Below I listed the offending call
stacks in my IDE. This makes sense as the associated drop down JList
needs new model. However, it makes fixing JComboBox.setModel(null)
hard or impossible.
Exception in thread "main" java.lang.IllegalArgumentException: model
must be non null
at javax.swing.JList.setModel(JList.java:1674)
at
javax.swing.plaf.basic.BasicComboPopup$Handler.propertyChange(BasicComboPopup.java:939)
at
java.beans.PropertyChangeSupport.fire(PropertyChangeSupport.java:335)
at
java.beans.PropertyChangeSupport.firePropertyChange(PropertyChangeSupport.java:327)
at
java.beans.PropertyChangeSupport.firePropertyChange(PropertyChangeSupport.java:263)
at java.awt.Component.firePropertyChange(Component.java:8413)
at javax.swing.JComboBox.setModel(JComboBox.java:322)
2. JTextComponent.setComponent(null) can be fixed but code change in
BasicTextUI is also required.
Since setting null model to JComboBox, JList and JTextComponent is
impossible or dangerous, just as you mentioned, we could set a non
null new model to these UI components just for the purpose of having
the side effect of removing listeners from old model. Are you ok with
this approach?
Yes. Please, try this and run the html swing automated tests
from the test/javax/swing/text/html diroctory to check possible
regressions.
Thanks,
Alexandr.
By the way, I will investigate your another question "Could you also
check that the StyleConstants.ComponentAttribute property value can't
be rewritten by the JDK code or by public methods." and reply soon.
Best regards,
Frank
On 11/13/2012 7:49 PM, Alexander Scherbatiy wrote:
On 11/13/2012 11:53 AM, Frank Ding wrote:
Hi Alexandr,
As for your comment "Could you create an issue that a
JComboBox.setModel(null) call throws NPE? You could fix it before
the 7189299 issue. ", I created a bug with internal review ID of
2381499 on JComboBox.setModel(null). But test shows that
JPasswordField.setDocument(null), JTextField.setDocument(null),
JList.setModel(null) and JTextArea.setDocument(null) all throw
NPE. Particularly, JList.setModel(null) has null check and throws
IllegalArgumentException("model" must be non null"). Shall I
include those components as well?
There is the javadoc for the JList setModel() method: Throws
IllegalArgumentException - if model is null. So it is undesirable
to change the public API.
However, it is possible to set a new empty model for the JList.
The list listeners should be removed from the old model in this case.
You could have only 2 issues: one for components that allow to
set a null model but throws NPE (like JComboBox) and another for
components that does not allow to set null model but they do not
remove listeners from the old model in case if a new model is set.
Thanks,
Alexandr.
Thanks for your guidance in advance.
Best regards,
Frank
On 11/8/2012 10:55 PM, Alexander Scherbatiy wrote:
On 11/7/2012 10:56 AM, Frank Ding wrote:
Hi Alexandr,
Unfortunately, all the JComponents involved in
FormView.createComponent() have bugs!
I have done tests for all other swing components, i.e.
JCheckBox, JRadioButton, JPasswordField, JTextField, JList,
JComboBox and JTextField. Sadder news is that if we fix all
components in the same way as I did for JButton, we need to
subclass them all, resulting in JCheckBoxWrapper,
JRadioButtonWrapper, JPasswordFieldWrapper, JTextFieldWrapper,
JListWrapper, JComboBoxWrapper, JTextFieldWrapper plus
JButtonWrapper! This approach becomes unacceptable when all
swing components are affected.
Shall we fix it in other way illustrated below?
1. Whenever a swing component is created, it is kept in
AttributeSet, as what is now for model.
2. In FormView.createComponent(), the old swing component can be
retrieved via
attr.getAttribute(StyleConstants.ComponentAttribute). Note that
ComponentAttribute is newly added.
This way should be carefully investigated that it does not
introduce new memory leaks.
Could you also check that the
StyleConstants.ComponentAttribute property value can't be
rewritten by the JDK code or by public methods.
3. Before setting shared model to a newly initialized swing
component, call oldComp.setModel(null), delegating deregistration
to setModel method.
4. Seems some setModel such as AbstractButton.setModel() can
function well when the param, new model, is null while
JComboBox.setModel() will throw NPE in case of null param.
Could you create an issue that a JComboBox.setModel(null)
call throws NPE? You could fix it before the 7189299 issue.
Thanks,
Alexandr.
5. Add null check code in those problematic setModel or
setDocument methods.
Any idea is appreciated. Thanks.
Best regards,
Frank
On 11/6/2012 9:26 PM, Alexander Scherbatiy wrote:
On 11/5/2012 11:20 AM, Frank Ding wrote:
Hi Alexander,
Based on your comments last time, I refined my patch of v6
and offered v7 @
http://cr.openjdk.java.net/~dingxmin/7189299/webrev.07/
This version of the fix looks good for me.
It seems that it is the only good way to check that a
button model contains AbstarctButton.Handler listener.
Could you also check that others models used in the
createComponent() method do not have such memory leaks (even the
NPE is not thrown)?
4. Could you add the check that the action listener is
invoked only once after a component tree updating and the
action does the same that it does before a component tree
updating?
Answer: I am afraid I could not make it in the auto test
(bug7189299.java) but I can achieve it to some degree in manual
test FormSubmit, the one you illustrated below.
My idea of checking it in FormSubmit.java is subclassing
JEditorPane and overriding 'public EditorKit getEditorKit()'
where stack traces can be examined in the overridden method to
make sure FormView.submitData occurs only once. This approach
works because FormView.submitData() calls
JEditorPane.getEditorKit but is tricky. However, it's the only
way I can think of without any additional framework support.
If you are in favor of adding the check in FormSubmit.java, I
am willing to do that.
At least a separated manual test can be added that asks a
user to put a response.html file to a server and according to
the server url checks that JTeditor pane show the response text
after a button pressing.
html = new JEditorPane("text/html",
"<html><body><form action=\"" + userURL + "\">"
+ "<input type=submit name=submit
value=\"submit\"/>"
+ "</form></body></html>");
response.html:
<html> <body> Hello World! </body> </html>
Thanks,
Alexandr.
Thanks again for all your comments and your support. Once
again, if you have any further concern or comment, please don't
hesitate to let me know.
Best regards,
Frank
