Make Socket IO behave like MRI in JRuby 9k

[haberbyte]

There is an issue with the pow webserver that only occurs with JRuby (basecamp/pow#14).

It is caused in the nack rack server in the following spot: https://github.com/josh/nack/blob/master/lib/nack/server.rb#L142

With JRuby i get:

IOError: Socket is not connected
org/jruby/ext/socket/RubyBasicSocket.java:423:in `close_read'

Here the link to RubyBasicSocket.java:

jruby/core/src/main/java/org/jruby/ext/socket/RubyBasicSocket.java

Line 423 in c510778

shutdownInternal(context, 0);

I'm not sure what shutdownInternal(context, 0); really does and why it fails.

But i would love to see this fixed and i'm willing to help (if i am able to).
It would be great if JRuby 9k can address this.

What further information to debug this issue can i provide?

[headius]

shutdownInternal(context, 0) should shut down just the read channel of the socket.

Have you tried this on master? Is there an easy way for me to reproduce this?

[headius]

I just tried something I think should reproduce the issue, but it passes on master and JRuby 1.7.18. Maybe we're done? Maybe I don't have a good enough case?

us = UNIXServer.new('/tmp/foo')
uc = UNIXSocket.new('/tmp/foo')
uc.close_read

[haberbyte]

I'm afraid it's not a good enough case and it's a bit more complicated :-(

I had a hard time coming up with a small enough example to reproduce this issue, but i came up with something that might help.

You need to run two ruby scripts, a server.rb and a client.rb.

server.rb:

require 'fcntl'
require 'socket'

server = UNIXServer.new('/tmp/foo')
server.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC)

clients = []
buffers = {}

loop do
  listeners = clients
  listeners << server unless server.closed?

  readable, writable = nil
  begin
    readable, writable = IO.select(listeners, nil, nil, 60)
  rescue Errno::EBADF
  end

  next unless readable

  readable.each do |sock|
    if sock == server
      client = server.accept_nonblock
      client.fcntl(Fcntl::F_SETFD, Fcntl::FD_CLOEXEC)
      clients << client
    else
      client, buf = sock, buffers[sock] ||= ''

      begin
        buf << client.read_nonblock(1024)
      rescue EOFError
        sock.close_read

        buffers.delete(client)
        clients.delete(client)

        puts 'WORKING'
        exit
      end
    end
  end
end

client.rb:

require 'socket'

s = UNIXSocket.new("/tmp/foo")
s.send "hello", 0

Now in one terminal run ruby server.rb and in another one run ruby client.rb.

The output for MRI on terminal one:

$ ruby server.rb
WORKING

The output for JRuby on terminal one:

$ ruby server.rb
IOError: Socket is not connected
                close_read at org/jruby/ext/socket/RubyBasicSocket.java:451
  chained_2_rescue_line_31 at server.rb:33
                    (root) at server.rb:31
                      each at org/jruby/RubyArray.java:1613
                    (root) at server.rb:22
                      loop at org/jruby/RubyKernel.java:1507
                    (root) at server.rb:10

[headius]

The reproduction looks great, thank you. Will investigate.

[headius]

Five minutes later, I have a fix. Pushing shortly.

[headius]

My fix is on master and I'll be adding some tests to MRI shortly.

[haberbyte]

That's amazing... ❤️

Now that i see what the fix looks like i feel kinda bad that i didn't come up with it myself :-(
Sometimes it's just hard to get started...don't know much about the underlying implementation.

Anyway, thank you, you rock. Achievement unlocked 😉

[headius]

Tip for the future: if there's an unusual error, you can usually pass -Xbacktrace.style=raw to JRuby to get the real Java trace. From there it's often easy to figure it out :-)

Please let me know if there's anything else keeping pow from working on JRuby 9k.

[headius]

Tests added to MRI in ruby/ruby@4eacd5d.