Possible concurrency issue in Drools
by Norman C
Hi All,
I recently ran into an issue which I believe might point to a concurrency
issue. My server stopped processing new requests, so I did a thread dump. In
examining the dump, I found that all of the processing threads, save two, were
blocking while trying to acquire the lock in NamedEntryPoint.insert. Both of
the other two threads appeared to be infinitely looping in the
NamedEntryPoint.insert method. Here are snippets of the stack traces:
ActiveMQ Session Task" prio=10 tid=0x00002aab0003b000 nid=0x7b98 runnable
[0x000000004c086000..0x000000004c087c90] java.lang.Thread.State: RUNNABLE
at org.drools.util.ObjectHashMap.remove(ObjectHashMap.java:121) at
org.drools.common.SingleThreadedObjectStore.removeHandle(SingleThreadedObjectStore.java:150)
at org.drools.common.NamedEntryPoint.retract(NamedEntryPoint.java:296)
at org.drools.common.NamedEntryPoint.retract(NamedEntryPoint.java:245) at
org.drools.reteoo.ReteooWorkingMemory$WorkingMemoryReteExpireAction.execute(ReteooWorkingMemory.java:350)
at
org.drools.common.AbstractWorkingMemory.executeQueuedActions(AbstractWorkingMemory.java:1488)
at org.drools.common.NamedEntryPoint.insert(NamedEntryPoint.java:158)
at org.drools.common.NamedEntryPoint.insert(NamedEntryPoint.java:122) at
org.drools.common.NamedEntryPoint.insert(NamedEntryPoint.java:80) at
org.drools.common.NamedEntryPoint.insert(NamedEntryPoint.java:28) at
ActiveMQ Session Task" prio=10 tid=0x000000005a35cc00 nid=0xdf6 runnable
[0x000000004a268000..0x000000004a269a90] java.lang.Thread.State: RUNNABLE
at org.drools.util.AbstractHashTable.resize(AbstractHashTable.java:115) at
org.drools.util.ObjectHashMap.put(ObjectHashMap.java:78) at
org.drools.common.SingleThreadedObjectStore.addHandle(SingleThreadedObjectStore.java:136)
at org.drools.common.NamedEntryPoint.insert(NamedEntryPoint.java:113)
at org.drools.common.NamedEntryPoint.insert(NamedEntryPoint.java:80) at
org.drools.common.NamedEntryPoint.insert(NamedEntryPoint.java:28) at
So it seems like one while the first thread is holding the lock and is
attempting to remove an object handle from the object store in NamedEntryPoint,
the other thread is trying to resize that same object store in response to an
addHandle call that puts it over the threshold. I haven't worked out exactly
how these concurrent accesses to the same object store by two different threads
causes an infinite loop in both threads, but it seems like the call to
SingleThreadedObjectStore.addHandle should be preceded by acquiring the lock.
Is this correct? I can imagine that resizing a large hash map could potentially
take a long time and thus synchronizing this call could impact performance, but
somehow, the action of resizing the table must be protected in some way from
adversely impacting other operations on the table.
Any help would be appreciated.
thanks,
Norman
15 years, 7 months
- 1
- 0