On Linux, ServerSocket.accept() throws "java.net.SocketTimeoutException: Accept timed out" if the call is interrupted by a signal.

To see the error, run the following program:

public class SocketAcceptTest {
    public static void main(String[] args) throws Exception {
        new ServerSocket(0).accept();
    }
}

Then find tid of the main thread, and send a signal to it, e.g. SIGPIPE which is caught but ignored by the JVM:

kill -SIGPIPE <tid>

Usually <tid> = <pid> + 1

The program will terminate with SocketTimeoutException, although the operation should never time out.

Exception in thread "main" java.net.SocketTimeoutException: Accept timed out
        at java.base/java.net.PlainSocketImpl.socketAccept(Native Method)
        at java.base/java.net.AbstractPlainSocketImpl.accept(AbstractPlainSocketImpl.java:458)
        at java.base/java.net.ServerSocket.implAccept(ServerSocket.java:565)
        at java.base/java.net.ServerSocket.accept(ServerSocket.java:533)
        at Test.main(Test.java:6)

The problem always happens on JDK 11 and also on JDK 13/14, if run with -Djdk.net.usePlainSocketImpl=true

Works fine on JDK 8 and on JDK 13/14, if run with new Socket implementation (JEP 353).

The original problem was encountered when profiling Tomcat with async-profiler, see https://github.com/jvm-profiling-tools/async-profiler/issues/295

Evaluation:

PlainSocketImpl_socketAccept() calls NET_Timeout (i.e. poll) with a negative timeout, which should mean "wait infinitely". However, a signal may interrupt a call, it returns EINTR, and then the loop in NET_Timeout immediately breaks because of the negative value:

        if (rv < 0 && errno == EINTR) {
            jlong newNanoTime = JVM_NanoTime(env, 0);
            nanoTimeout -= newNanoTime - prevNanoTime;
            if (nanoTimeout < NET_NSEC_PER_MSEC) {
                return 0;
            }
            prevNanoTime = newNanoTime;
        }