Errno::EBADF is sometimes raised instead of IOError when TCPSocket#readline is called after TCPSocket#close

When TCPSocket#readline is called after a TCPSocket#close on the same client socket, if the #close comes from another thread, sometimes an Errno::EBADF is the result instead of an IOError: closed stream.

Tracing through the code, what's happening is that the ChannelDescriptor responds to the getline() call with a BadDescriptorException because the underlying Channel responds false to isOpen(), which on the face of it seems wrong but I'm not familiar enough with the semantics to make a convincing argument. Also, given that the readStream.getline() call is guarded by a beforeBlockingCall(), I suspect that this actually indicates a race somewhere else in the code.

It is proving difficult to narrow this down to a single test case, but the following code at least demonstrates the problem:

  require 'socket'

  server = TCPServer.new(0)
  server_thread = Thread.new{ server.accept while true }

  exceptions = []

  1000.times do |i|
    client_socket = TCPSocket.new("localhost", server.addr[1])
    close_thread = Thread.new{ sleep(Kernel.rand/100); client_socket.close }
    read_thread = Thread.new do 
      sleep(Kernel.rand/100); 
      begin
        client_socket.readline
      rescue Exception => e
        exceptions << e
      end
    end
  
    close_thread.join
    read_thread.join
  end

  p exceptions.map{|e| e.class.to_s}.inject({}){|m,r| m.merge r => true}.keys.sort

The final output is, for me:

["EOFError", "Errno::EBADF", "Errno::EPIPE", "IOError"]

EOFError and IOError are expected. EBADF and EPIPE are not.