NativeEventThread's run() method's infinite loop was implemented. The
loop looks like this:
while (null != this.browserControlCanvas) {
synchronized (this.browserControlCanvas.getTreeLock()) {
nativeProcessEvents(nativeWebShell);
if (null != listenersToAdd && !listenersToAdd.isEmpty()) {
tempEnum = listenersToAdd.elements();
while (tempEnum.hasMoreElements()) {
nativeAddListener(nativeWebShell,
(WebclientEventListener)
tempEnum.nextElement());
}
listenersToAdd.clear();
}
}
}
The problem I was observing was that
nativeProcessEvents(nativeWebShell) would crash due to the fact that
the nativeWebShell, which is actually an WebShellInitContext instance,
had been de-allocated. This de-allocation happens as a result of the
WindowControlImpl.delete() method, which looks like this:
public void delete()
{
Assert.assert(null != eventThread, "eventThread shouldn't be null at delete time");
eventThread.delete();
eventThread = null;
nativeDestroyInitContext(nativeWebShell);
nativeWebShell = -1;
}
nativeDestroyInitContext de-allocates the WebShellInitContextInstance.
You can see that the first thing done is to delete the eventThread().
NativeEventThread.delete() looks like this:
public void delete()
{
// setting this to null causes the run thread to exit
synchronized(this.browserControlCanvas.getTreeLock()) {
browserControlCanvas = null;
}
...
}
If you compare NativeEventThread.delete() with the infinite loop in
NativeEventThread.run(), you'll see that the fact that they both
synchronize on the same object doesn't protect us from the following
case:
NativeEventThread: The infinite loop checks to see if the
browserControlCanvas is null, then does synchronize on
browserControlCanvas.getTreeLock(), then calls processNativeEvents().
meanwhile
WindowControlImpl thread: delete() calls NativeEventThread.delete(),
which does synchronize on browserControlCanvas.getTreeLock().
During NativeEventThread.delete(), synchronized section,
browserControlCanvas is set to null.
NativeEventThread: because the check for null browserControlCanvas
occurrs outside of the synchronized block, it's not recheked, and
thus, the event loop continues to process when it shouldn't.
The fix is to change the event loop to look like this:
while (true) {
synchronized (this.browserControlCanvas.getTreeLock()) {
// this has to be inside the synchronized block!
if (null == this.browserControlCanvas) {
return;
}
nativeProcessEvents(nativeWebShell);
if (null != listenersToAdd && !listenersToAdd.isEmpty()) {
tempEnum = listenersToAdd.elements();
while (tempEnum.hasMoreElements()) {
nativeAddListener(nativeWebShell,
(WebclientEventListener)
tempEnum.nextElement());
}
listenersToAdd.clear();
}
}
}
- reduces a number of c++<--> java calls
- added NULL checks
- made DOMAccessor to be secure
- added util and tests packages
- wrote test applets
- updated README
a=edburns
bug=32011
This change enables the current webclient API to be called from native
code.
It adds makefile and conditional compilation logic.
If the user defines BAL_INTERFACE in their environment before building
webclient, -DBAL_INTERFACE is added to LCFLAGS. This causes code in
jni_util_export.cpp to behave differently due to the conditional
compilation logic.
I've broken out the 8 functions that are necessary to call into the
Webclient JNI methods into jni_util_export.{h,cpp}.
I've created a new pair of files, bal_util.{h,cpp} that contain function
declarations and definitions that are used when src_moz is built with
BAL_INTERFACE. bal_util.obj is not built, nor added to webclient.dll if
building without BAL_INTERFACE.
See the page
http://www.mozilla.org/projects/blackwood/webclient/design/uno-transition.html
for a design document description of these changes.
r=ashuk
a=edburns
This set of changes replaces all occurrences of
env->Func(args...)
with
::util_Func(env, args...)
Except of course, for the implementations of the above mentioned
::util_Func() functions.
This is done to allow the JNI functions to be called from a non JNI
context, such as UNO.
also modifies EmbeddedMozilla so this code is exercised.
I have changed EmbeddedMozilla to be a stub-like class that simply
displays a Frame with a single Button, titled "New Window". Pressing
this button causes an EMWindow to be created and displayed. EMWindow is
basically the former EmbeddedMozilla renamed, with modifications to the
WindowListener implementation to call the BrowserControl deallocation
method.
I've added a delete() method to ImplObect:
* I know Java has automatic garbage collection and all, but explicitly
* adding a delete method helps the gc algorithm out. <P>
* Subclasses should override this and call super.delete() at the end of
* their overridden delete() method.
* @see org.mozilla.webclient.wrapper_native.ImplObjectNative#delete
and ImplObjectNative:
* Note how we call super.delete() at the end. THIS IS VERY IMPORTANT. <P>
* Also, note how we don't de-allocate nativeWebShell, that is done in
* the class that owns the nativeWebShell reference, WindowControlImpl.
* ImplObjectNative subclasses that further override delete() are <P>
<CODE><PRE>
BookmarksImpl.java
EventRegistrationImpl.java
NativeEventThread.java
WindowControlImpl.java
</PRE><CODE> <P>
* All other ImplObject subclasses don't have any local Ivars and thus
* don't need to override delete().
I've added a delete() method to BrowserControlImpl:
* Called from BrowserControlFactory.deleteBrowserControl() <P>
* The order of deletion of objects is very important! <P>
* We don't allow deletion if the Canvas is showing. <P>
In BrowserControlImpl's delete(), the important delete()s is for
WindowControlImpl:
* First, we delete our eventThread, which causes the eventThread to
* stop running. Then we call nativeDestroyInitContext(), which
* deallocates native resources for this window.
As stated above, NativeEventThread.delete() is called:
* This is a very delicate method, and possibly subject to race
* condition problems. To combat this, our first step is to set our
* browserControlCanvas to null, within a synchronized block which
* synchronizes on the same object used in the run() method's event
* loop. By setting the browserControlCanvas ivar to null, we cause the
* run method to return.
After all of this deleting, we return from
BrowserControlFactory.delete().
edburns%acm.org
rpallath%eng.sun.com
sdv%sparc.spb.su
ashuk%eng.sun.com
idk%eng.sun.com
sherry.shen%sun.com