Hi Mandy,
This seems quite an involved solution compared to what is described in
the CR. The lazy initialization of the bcp seems to be a main part of
the problem (or rather the requisite synchronization thereof). Do we
really need to lazily initialize the bcp?
Even if we do, can we not simply avoid the nested locking by only
locking when needing to initialize the bcp (and cache it in a volatile)
and move the getProperty outside of that synchronized section. No nested
locking so no deadlocks.
Cheers,
David
Mandy Chung said the following on 09/17/10 09:30:
Fix 6977738: Deadlock between java.lang.ClassLoader and
java.util.Properties
Webrev at:
http://cr.openjdk.java.net/~mchung/6977738/webrev.00/
ClassLoader.getResource locking the system properties object is
deadlock prone. For example, the sun.misc.Launcher.getBootstrapClassPath()
method is a synchronized method. A thread calling
ClassLoader.getResource method attempts to synchronize on
the system properties object while holding sun.misc.Launcher.class.
It will deadlock if there is another thread is holding the monitor
lock of the system properties object (e.g. calls the Properties.save()
method) while it attempts to load a resource file. In addition,
ClassLoader.getResource may invoke the ZipFile initialization that
also needs to get the system property.
This fix caches the system properties at initialization so that
the ZipFile and Launcher will read the system property from the
private Properties object. In addition, I make some minor
changes in Properties.save and storeToXML method to confine the
synchronization needed and also have the DownloadManager to
maintain its bootclasspath.
Thanks
Mandy
