[GR-45893] jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE not available in native-image application

[beckermarc]

I am facing issues using jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE in native images. The following code runs fine on a JVM, but fails in a native image:

AsynchronousSocketChannel socket = AsynchronousSocketChannel.open();
socket.setOption(jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE, 100);

Error in native-image:

Exception in thread "main" java.lang.UnsupportedOperationException: 'TCP_KEEPIDLE' not supported
        at java.base@17.0.6/sun.nio.ch.AsynchronousSocketChannelImpl.setOption(AsynchronousSocketChannelImpl.java:461)
        at com.example.repro.Application.main(Application.java:15)

The class jdk.net.ExtendedSocketOptions registers some platform-dependent optionally available socket options. All of this is orchestrated by sun.net.ext.ExtendedSocketOptions, which dynamically loads jdk.net.ExtendedSocketOptions if available. jdk.net.ExtendedSocketOptions then performs initialization of these additional options in a static code block.

For some reason, this static code block is not properly executed in the native-image. I know about build-time initialization of classes in native-images, but do not fully understand if this has an impact here.

To enable dynamic lookup of jdk.net.ExtendedSocketOptions I also put the following information into a reflect-config.json:

[
    {
        "name": "jdk.net.ExtendedSocketOptions"
    }
]

Below I have provided a simple application to reproduce the issue. There I have provided additional debug output, which checks the conditions that check some boundary conditions for the exception to occur:

AsynchronousSocketChannel socket = AsynchronousSocketChannel.open();
Set<SocketOption<?>> options = socket.supportedOptions();
System.out.println("Has TCP_KEEPIDLE? " + options.contains(jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE));
System.out.println("Options: " + options);

Interestingly this produces the following output in a native-image:

Has TCP_KEEPIDLE? false
Options: [TCP_NODELAY, TCP_KEEPCOUNT, TCP_KEEPINTERVAL, SO_REUSEADDR, SO_SNDBUF, SO_RCVBUF, TCP_KEEPIDLE, SO_INCOMING_NAPI_ID, TCP_QUICKACK, SO_KEEPALIVE, SO_REUSEPORT]

Notice that the contains check evaluates to false, however the stringified options DO INCLUDE TCP_KEEPIDLE. Maybe this helps in finding the issue. My guess is this has something to do with build-time class initialization and code optimization, but I don't fully understand this behaviour.

Steps to reproduce the issue
Please include both build steps as well as run steps

1. Unpack the application: tcp-keepidle-repro.zip
2. mvn native:compile
3. java -jar target/tcp-keepaliveidle-0.0.1-SNAPSHOT.jar -> works
4. target/tcp-keepaliveidle -> fails

Describe GraalVM and your environment:

• GraalVM version: OpenJDK 64-Bit Server VM GraalVM CE 17.0.7-dev+4.1 (build 17.0.7+4-jvmci-23.0-b10, mixed mode, sharing)
• JDK major version: 17
• OS: Ubuntu 22.04
• Architecture: AMD64

[beckermarc]

I added some more output to my repro application, which better explains the current behaviour and points more to a bug:

AsynchronousSocketChannel socket = AsynchronousSocketChannel.open();
Set<SocketOption<?>> options = socket.supportedOptions();

System.out.println("Has TCP_KEEPIDLE (obj)? " + options.contains(jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE));
System.out.println("Has TCP_KEEPIDLE (string)? " + options.stream().anyMatch(s -> s.name().equals(jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE.name())));
System.out.println("Options: " + options);

SocketOption<?> so = options.stream().filter(s -> s.name().equals(jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE.name())).findFirst().get();
System.out.println("Constant identity: " + System.identityHashCode(jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE));
System.out.println("Set identity: " + System.identityHashCode(so));
socket.setOption(jdk.net.ExtendedSocketOptions.TCP_KEEPIDLE, 100);

The code above produces the following output when run on the JVM:

Has TCP_KEEPIDLE (obj)? true
Has TCP_KEEPIDLE (string)? true
Options: [TCP_KEEPCOUNT, TCP_QUICKACK, SO_KEEPALIVE, SO_REUSEADDR, SO_INCOMING_NAPI_ID, TCP_KEEPIDLE, SO_SNDBUF, SO_RCVBUF, SO_REUSEPORT, TCP_NODELAY, TCP_KEEPINTERVAL]
Constant identity: 186276003
Set identity: 186276003

In a native-image it produces the following output:

Has TCP_KEEPIDLE (obj)? false
Has TCP_KEEPIDLE (string)? true
Options: [SO_KEEPALIVE, TCP_NODELAY, TCP_KEEPINTERVAL, SO_REUSEADDR, SO_REUSEPORT, TCP_QUICKACK, SO_RCVBUF, TCP_KEEPCOUNT, SO_SNDBUF, SO_INCOMING_NAPI_ID, TCP_KEEPIDLE]
Constant identity: 1710198919
Set identity: 95685867
Exception in thread "main" java.lang.UnsupportedOperationException: 'TCP_KEEPIDLE' not supported
        at java.base@17.0.6/sun.nio.ch.AsynchronousSocketChannelImpl.setOption(AsynchronousSocketChannelImpl.java:461)
        at com.example.repro.Application.main(Application.java:29)

As you can see there exist two object instances for TCP_KEEPIDLE in the native-image. However the JDK source code relies on object identity comparison (the relevant class doesn't implement hashCode and equals) and uses the public static final SocketOption<Integer> TCP_KEEPINTERVAL declared object everywhere. Can you explain why there exist two different instances in the native-image?

[oubidar-Abderrahim]

Thank you for reporting this, This is tracked internally on GR 45893

[beckermarc]

Hey, just wanted to ask if there is any update on this issue?
Have you been able to identify any clues what is causing this issue (and maybe potential workarounds for the time being)?
Thanks!

[fniephaus]

Can you explain why there exist two different instances in the native-image?

Because one instance was created at build-time, and the other at run-time. More specifically, it is sun.nio.ch.AsynchronousSocketChannelImpl$DefaultOptionsHolder that is initialized at build-time and then incorrectly persisted as part of the executable. b6654d0 makes sure its class initializer runs again at run-time, so that the set is reinitialized with the correct values. With my patch, I now get:

tcp-keepidle-repro$ ./target/tcp-keepaliveidle 
Is Linux? true
Has TCP_KEEPIDLE? true
Options: [TCP_QUICKACK, SO_RCVBUF, SO_KEEPALIVE, TCP_NODELAY, TCP_KEEPCOUNT, SO_SNDBUF, TCP_KEEPIDLE, SO_REUSEPORT, SO_INCOMING_NAPI_ID, TCP_KEEPINTERVAL, SO_REUSEADDR]
Constant identity: 164063280
Set identity: 164063280

If you want to try this out now, you need to build Native Image from source with my patch. Soon after #7440 is merged, the latest dev build will ship the fix.

Hope this helps!

-- Fabio

[beckermarc]

Awesome! Thanks for looking into this :)
Can this reinitialization also be configured somehow in native-image.properties, as a workaround for now?

[fniephaus]

Yes, I believe you can use --initialize-at-run-time=sun.nio.ch.AsynchronousSocketChannelImpl$DefaultOptionsHolder. You can add the flag via your pom.xml file like so:

<buildArgs>
	<buildArg>--initialize-at-run-time=sun.nio.ch.AsynchronousSocketChannelImpl$DefaultOptionsHolder</buildArg>
</buildArgs>