From: Bill Kelly <billk@...>
Date: 2006-03-29T06:11:31+09:00
Subject: Re: timeout when listening with TCPServer

From: "Joel VanderWerf" <vjoel@path.berkeley.edu>
>
> I think it might be clearer to write it as
> 
> session = Timeout.timeout 2 do
>  server.accept
> end
> 
> Thread.start(session) do |s| new_connection s end

I think I'd do it that way too.  (Probably adding a 
rescue for Timeout::Error.)

I wonder if, there's still any possible race condition
in this version.

  session = Timeout.timeout 2 do
    server.accept
    # If timeout fires "here" would the socket be leaked?
    # (eventual probable reaping by GC notwithstanding)
  end

Not that I would lose much sleep over it.  But I try
to write code without any possibility of race conditions
whenever possible.

I'm guessing the race condition does exist there.  Anyone
know whether assignments are atomic with regard to
exceptions?  E.g.

  begin
    Timeout.timeout 2 do
      session = server.accept
      #      ^^^ atomic?
    end
  rescue Timeout::Error
  end

Any possibility of the timeout interrupting the assignment
to "session" there?  If assignment is atomic then we
should be able to avoid leaking the socket.

(Just wondering... not too worried about it in practical
terms with this particular example.)


Regards,

Bill