[
https://issues.apache.org/jira/browse/FELIX-2597?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Richard S. Hall updated FELIX-2597:
-----------------------------------
Description:
When the framework resolves a bundle, it fires the RESOLVED event to
synchronous bundle listeners while holding the global lock. This caused a
deadlock in a specific scenario where two threads T1 and T2 were racing to
resolve bundles B1 and B2, respectively. T1 was resolving a dynamic import for
B1 which started from a loadClassInternal() call from the JVM which locks B1's
class loader. T1 lost the race for the global lock and had to wait for T2 to do
its resolve of B2. T2 resolved B2 and ended up firing a RESOLVED event to B1.
In B1's callback method, a loadClassInternal() was triggered by T2 when it
accessed some static, which caused it to deadlock when the JRE tried to lock
the B1's class loader on T2, which was already held by T1.
This issue is actually related to the JVM being too aggressively in its locking
of class loaders:
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=4670071
Regardless, we might be able to loosen this up a little bit by not holding the
global lock at all. This might slightly violate the definition of synchronous
bundle listener, since the state could change, but at least it will still be
delivered synchronously on the instigating thread.
More information here:
https://glassfish.dev.java.net/issues/show_bug.cgi?id=13329
was:
When the framework resolves a bundle, it fires the RESOLVED event to
synchronous bundle listeners while holding the global lock. This caused a
deadlock in a specific scenario where two threads T1 and T2 were racing to
resolve bundled B1 and B2, respectively. T1 was resolving a dynamic import for
B2 which started from a loadClassInternal() call from the JVM which locks B2's
class loader. T1 lost the race for the global lock and had to wait for T2 to do
its resolve of B2. T2 ended up firing a RESOLVED event to B1. In B1's callback
method, a loadClassInternal() was triggered by T2 when it accessed some static,
which caused it to deadlock when the JRE tried to lock the B1's class loader on
T2, which was already held by T1.
This issue is actually related to the JVM being too aggressively in its locking
of class loaders:
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=4670071
Regardless, we might be able to loosen this up a little bit by not holding the
global lock at all. This might slightly violate the definition of synchronous
bundle listener, since the state could change, but at least it will still be
delivered synchronously on the instigating thread.
More information here:
https://glassfish.dev.java.net/issues/show_bug.cgi?id=13329
> Deadlock during delivery of resolved event
> ------------------------------------------
>
> Key: FELIX-2597
> URL: https://issues.apache.org/jira/browse/FELIX-2597
> Project: Felix
> Issue Type: Improvement
> Components: Framework
> Affects Versions: framework-3.0.2
> Reporter: Richard S. Hall
> Assignee: Richard S. Hall
> Priority: Minor
> Fix For: framework-3.2.0
>
>
> When the framework resolves a bundle, it fires the RESOLVED event to
> synchronous bundle listeners while holding the global lock. This caused a
> deadlock in a specific scenario where two threads T1 and T2 were racing to
> resolve bundles B1 and B2, respectively. T1 was resolving a dynamic import
> for B1 which started from a loadClassInternal() call from the JVM which locks
> B1's class loader. T1 lost the race for the global lock and had to wait for
> T2 to do its resolve of B2. T2 resolved B2 and ended up firing a RESOLVED
> event to B1. In B1's callback method, a loadClassInternal() was triggered by
> T2 when it accessed some static, which caused it to deadlock when the JRE
> tried to lock the B1's class loader on T2, which was already held by T1.
> This issue is actually related to the JVM being too aggressively in its
> locking of class loaders:
> http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=4670071
> Regardless, we might be able to loosen this up a little bit by not holding
> the global lock at all. This might slightly violate the definition of
> synchronous bundle listener, since the state could change, but at least it
> will still be delivered synchronously on the instigating thread.
> More information here:
> https://glassfish.dev.java.net/issues/show_bug.cgi?id=13329
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
