&mut -> & of session in event loop

[stepancheg]

To process session from multiple threads, we need to use the session as &session instead of &mut session, so the session can be put into Arc or something.

This PR extracts event_loop function to handle all requests except init and destroy, and handles FUSE_DESTROY outside of event loop.

[cberner]

Can you say more about how you're planning to implement the multi-threading? I haven't looked into it very much, but this PR: #421 adds an op to clone the /dev/fuse fd and then create multiple sessions.

[stepancheg]

My idea was this: you pass a parameter to session creation, number of threads.

Then run function is modified like this, pseudocode:

fn run(self) {
  handshake();

  let this = Arc::new(self);

  let threads = Vec::new();
  for thread in 0..n_threads {
    let fd = clone_fd();
    let this = this.clone();
    threads.push(thread::spawn(|| this.event_loop());
  }

  join_threads();

  destroy();
}

So that fuser library manages threads.

That PR exposes low level operation Session::from_fd_initialized instead, which is less convenient, but also creates other problems, for example, Session::from_fd_initialized does not now know what features were negotiated. In particular, that PR incorrectly initializes Session.proto_version field. And it does call from_fd_initialized before initialization even happens (comment "Skip INIT - caller guarantees mount is initialized" is false when using provided test as reference). Also that PR contains changes unrelated to the problem. It defines FUSE_DEV_IOC_CLONE constant twice instead of using ioctl macro from nix. Comments in that PR are excessive, e.g. no need to have a comment "open a new /dev/fuse" literally before the call open("/dev/fuse"). Tests are written with sleeps so they will be flaky instead of proper synchronization. Overall, I don't like quality of that PR.

That said, exposing low level parts of working with FUSE may have its use.

[cberner]

Cool, thanks. Ya, that outline looks good to me