Richard Zowalla created OPENNLP-1976:
----------------------------------------
Summary: AbstractDL.close() does not wait for in-flight inference,
so a concurrent close can free the native session during a run
Key: OPENNLP-1976
URL: https://issues.apache.org/jira/browse/OPENNLP-1976
Project: OpenNLP
Issue Type: Bug
Reporter: Richard Zowalla
Fix For: 3.0.0
{{AbstractDL.close\(\)}} only guards against a second close \({{AtomicBoolean}}
compare\-and\-set\) and then calls {{OrtSession.close\(\)}}. It does not wait
for {{session.run}} calls that are still executing in {{SentenceVectorsDL}},
{{NameFinderDL}} or {{DocumentCategorizerDL}}. In ONNX Runtime Java,
{{OrtSession.closed}} is a plain boolean checked before the native call, so a
run that overlaps {{close\(\)}} can execute on a freed native session. The
result is undefined native behavior, up to a JVM crash, rather than an
exception. This happens in practice when an application server undeploys an
application while requests are still being served.
Suggested fix: guard every {{session.run}} with a read lock \(or an in\-flight
counter\) in a protected helper of {{AbstractDL}}, take the write lock in
{{close\(\)}}, and throw {{IllegalStateException}} for calls after close. If
in\-flight runs do not finish within a bounded time, log and skip freeing the
session instead of freeing it under a running call.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
