Hi, I'm new here. I'm wondering if folks will accept a contribution to fix https://issues.apache.org/jira/browse/HDFS-13585 ?
I can repro the issue by creating a Java application that calls hdfsConnectNewInstance: https://gist.github.com/emancebo/e413b0f53e4005c80cbf410fdb4c9cb9 The crash occurs in hdfsThreadDestructor. It seems that any of the calls on the JNIEnv can cause a crash. I think in this scenario, the JVM (there is only one - the one created for the Java program) is getting cleaned up as the application is exiting, hence it is unsafe to call any of the methods on JNIEnv in the destructor function. My idea is to add a check in getGlobalJNIEnv to see if we are already attached to a JVM (by calling GetEnv). If we're already attached, we don't need to call AttachCurrentThread, and we can also skip the block of code in hdfsThreadDestructor that calls DetachCurrentThread. We could add a flag in ThreadLocalState to indicate whether we need to perform the detach. Definitely open to other approaches though. Thanks in advance - Ed
