The first libhdfs operation that is performed is not thread-safe; this is because the creation of a JVM is not protected by a mutex.
We have been able to trigger this by doing the following:
1) Start a few GNOME sessions on the box. Make sure you are running the gnome volume manager. The volume manager will perform a GETATTR operation on any newly mounted file system.
2) Start fuse-dfs in debug mode. As soon as it starts, you will see two or more GETATTR calls almost instantly.
3) fuse-dfs segfaults; if you ran this with GDB, you'll see a stack trace coming from libhdfs starting up a new JVM.
I imagine you could replicate this more simply by having two threads that call libhdfs simultaneously.
I have a patch for fuse-dfs which avoids the problem, but we probably need to fix it in libhdfs itself.