dd_oome_notifier.sh and dd_crash_uploader.sh fail when JDK_JAVA_OPTIONS contains JMX or other port-binding flags

[florianmutter]

Tracer Version(s)

1.55.0

Java Version(s)

21.0.9

JVM Vendor

Eclipse Adoptium / Temurin

Bug Report

The dd_oome_notifier.sh (and dd_crash_uploader.sh) scripts spawned via -XX:OnOutOfMemoryError do not unset JDK_JAVA_OPTIONS, JAVA_TOOL_OPTIONS, or _JAVA_OPTIONS before launching a child java process. The child JVM therefore inherits the full application JVM configuration, which causes three distinct problems:

1. Port conflicts — flags like JMX remote (-Dcom.sun.management.jmxremote.port=9012) cause BindException because the parent JVM still holds the port
2. cgroup OOMKill — memory flags like -Xms/-Xmx or -XX:MaxRAMPercentage=90 cause the child JVM to compete with the still-alive parent for container memory, potentially triggering a kernel OOMKill
3. Lost OOM diagnostics — when the script fails for any of the above reasons, no OOME event reaches Datadog, and the original OOM exception details (stack trace, thread name) are also lost because -XX:+ExitOnOutOfMemoryError force-terminates after the handler runs

Actual output:

# java.lang.OutOfMemoryError: Metaspace
# -XX:OnOutOfMemoryError="/tmp/datadog/java/dd_oome_notifier.sh %p"
#   Executing /bin/sh -c "/tmp/datadog/java/dd_oome_notifier.sh 1"...
Agent Jar: /opt/datadog/apm/library/java/dd-java-agent.jar
Tags: host:order-664fc65797-2bclc,...
JAVA_HOME: /opt/java/openjdk
PID: 1
NOTE: Picked up JDK_JAVA_OPTIONS: -XX:MaxGCPauseMillis=4000 -XX:MinRAMPercentage=25 -XX:MaxRAMPercentage=90 -XX:MaxMetaspaceSize=128m -Dcom.sun.management.jmxremote -Dcom.sun.management.jmxremote.port=9012 -Dcom.sun.management.jmxremote.rmi.port=9012 ...
Picked up JAVA_TOOL_OPTIONS: -javaagent:/opt/datadog/apm/library/java/dd-java-agent.jar ...

Caused by: java.rmi.server.ExportException: Port already in use: 9012; nested exception is:
	java.net.BindException: Address already in use
	...
Error: Failed to generate OOME event
Terminating due to java.lang.OutOfMemoryError: Metaspace

Expected Behavior

The OOME event should be sent to Datadog successfully, regardless of what JDK_JAVA_OPTIONS or JAVA_TOOL_OPTIONS contain. When the script fails, the original OOM exception details (stack trace, thread name) should still be visible in the application logs.

Reproduction Code

1. Configure a JVM application with JMX remote monitoring via JDK_JAVA_OPTIONS:

   JDK_JAVA_OPTIONS=-Dcom.sun.management.jmxremote -Dcom.sun.management.jmxremote.port=9012 -Dcom.sun.management.jmxremote.rmi.port=9012 ...
   

2. Have dd-java-agent injected (e.g. via Admission Controller), which sets:

   JAVA_TOOL_OPTIONS=-javaagent:/opt/datadog/apm/library/java/dd-java-agent.jar -XX:OnOutOfMemoryError="/tmp/datadog/java/dd_oome_notifier.sh %p"
   

3. Trigger an OutOfMemoryError (in our case: java.lang.OutOfMemoryError: Metaspace)

4. The JVM invokes dd_oome_notifier.sh, which spawns a child java process:

   "$config_java_home/bin/java" -Ddd.dogstatsd.start-delay=0 -jar "$config_agent" sendOomeEvent "$config_tags"

5. This child process inherits JDK_JAVA_OPTIONS (including JMX port flags) and JAVA_TOOL_OPTIONS (including the agent jar) from the environment.

6. The child JVM tries to bind JMX to port 9012, which is still held by the dying parent JVM → BindException: Address already in use → "Error: Failed to generate OOME event"

[florianmutter]

Maybe relevant: we use the admission controller to inject the agent.

[jbachorik]

Hi @florianmutter - thanks for the report!
Will take a look at the fix shortly.

[florianmutter]

One more note: Setting DD_CRASHTRACKING_ENABLED=false or JVM Option -Ddd.crashtracking.enabled=false seem to not prevent the -XX:OnOutOfMemoryError and related options from being added.

[jbachorik]

There is a bit of a problem - the -XX:OnOutOfMemoryError is injected by the admission controller unconditionally. It does not really know that -Ddd.crashtracking.enabled=false was provided.
But we need to create the scripts regardless, because the JVM is already set up with paths to the scripts and you would get error logs if the scripts are missing.