TCPServer.for_fd raises IOError: closed stream

[dekellum]

I originally encountered this and reported it as puma/puma#1367. I'm reporting it here because its a jruby-specific issue and I intend to provide a minimized reproduction outside of Puma.

I find other references to for_fd() possibly being implemented on jruby: #940, #3494 Only for UNIX Server/Socket? It is possible to just do the same thing for TCP Server/Socket?

Minimum test ruby script is below. This can be easily tested on a Linux system with systemd and its testing command systemd-socket-activate, then using nc for the connecting client. These commands should be available on all recent mainstream Linux distributions including Debian (8+), Ubuntu (16+), Fedora, Arch, etc.

activate.rb

require 'socket'

sock = TCPServer.for_fd(3)

port, addr = Socket.unpack_sockaddr_in(sock.getsockname)

puts "Activated #{addr} #{port}"

client = sock.accept
client.puts "Hi! Goodbye!"
client.close

Run as follows (substitute jruby as needed):

% systemd-socket-activate -l 0.0.0.0:2000 ruby -v ./activate.rb

Run this in one terminal, then use nc in another terminal to connect to the socket. (You will need to ^C to disconnect.) This is shown below with a successful (MRI ruby) run:

% nc 127.0.0.1 2000
Hi! Goodbye!
^C

With MRI ruby the output on the server side is the following expected output:

% systemd-socket-activate -l 0.0.0.0:2000 ruby -v ./activate.rb
Listening on 0.0.0.0:2000 as 3.
Communication attempt on fd 3.
Execing ruby (ruby -v ./activate.rb)
ruby 2.2.7p470 (2017-03-28 revision 58194) [x86_64-linux]
Activated 0.0.0.0 2000

With jruby 9.1, an IOError is instead raised from the initial for_fd call:

% systemd-socket-activate -l 0.0.0.0:2000 jruby -v ./activate.rb
Listening on 0.0.0.0:2000 as 3.
Communication attempt on fd 3.
Execing jruby (jruby -v ./activate.rb)
jruby 9.1.12.0 (2.3.3) 2017-06-15 33c6439 OpenJDK 64-Bit Server VM 25.131-b11 on 1.8.0_131-b11 +jit [linux-x86_64]
IOError: closed stream
  for_fd at org/jruby/ext/socket/RubyBasicSocket.java:109
  <main> at ./activate.rb:5

[dekellum]

Just to confirm this isn't a recent regression, on jruby 1.7.27 the Error is different but raised in the same place:

systemd-socket-activate -l 0.0.0.0:2000 jruby17 -v ./activate.rb 
Listening on 0.0.0.0:2000 as 3.
Communication attempt on fd 3.
Execing jruby17 (jruby17 -v ./activate.rb)
jruby 1.7.27 (1.9.3p551) 2017-05-11 8cdb01a on OpenJDK 64-Bit Server VM 1.8.0_131-b11 +jit [linux-amd64]
Errno::EINVAL: Invalid argument - Invalid file
  for_fd at org/jruby/ext/socket/RubyBasicSocket.java:117
  (root) at ./activate.rb:5

[ahorek]

The problem is, that jruby is looking into a map of existing (registered) file descriptors, but if you use a descriptor that wasn't created inside jruby process it fails.

https://github.com/jruby/jruby/blob/master/core/src/main/java/org/jruby/ext/socket/RubyBasicSocket.java#L106

@kares @headius

[ahorek]

#5008

[headius]

Yes, this is not an easy one to fix. Our socket logic is based around JVM APIs, which means a TCPSocket always needs a Java TCPSocket underneath it. At the same time, we do try to carve out the real filedescriptor, usually without storing it somewhere. Then when you try to for_fd, we see a real FD, find no existing TCPSocket, and are unable to produce one.

This is another one that will be fixed by going with a fully-native socket API, likely in 9.2. See #4748 and others.

[headius]

This never happened for 9.2 and is certainly too big for a maintenance release.

[headius]

This would be fixed by supporting the FFI-based socket library, and is superseded by #6334. Until then it is not possible to resolve, since we can't create a TCPServer on the JDK from a raw file descriptor.