Buffered Socket#close may not return if TCP send buffer is full

[spuun]

Bug Report

In LavinMQ it's possible to force close client connections from the broker via the HTTP API. Today I stumbled over a bug where the HTTP request timed out, and after digging my conclusion is that it's because Socket#close never returned.

In this particular case it was a client that didn't process the messages LavinMQ sent to it, causing the TCP send buffer to fill up. The reason Socket#close doesn't return is because the write buffer in Socket contains data that #close wants to flush, which is impossible because the TCP buffer is full.

My first attempt was to use Buffered#unbuffered_close which as a first glance sounded like a solution, but Socket implement it as private. Instead i could use Socket#write_close.

PoC:

require "socket"
TCPServer.open(0) do |server|
  spawn { sleep if client = server.accept? }

  TCPSocket.open(server.local_address.address, server.local_address.port) do |client|
    client.sync = false

    ch = Channel(Nil).new
    spawn do
      loop do
        select
        when ch.receive
        when timeout 1.seconds
          # Ensure data in write buffer
          client.write "foo".to_slice if client.@out_count.zero?
          puts "Buffer full, closing socket"
          client.close
          # Never printed:
          puts "Socket closed"
        end
      end
    end

    data = Bytes.new(client.buffer_size // 2)
    loop do
      client.write data
      ch.send nil
    end
  end
end

Dunno if this should be treated as a bug, or a missing feature?

Maybe a #close(flush : Bool = true) is needed?

[ysbaddaden]

Also maybe a timeout on the flush itself wouldn't hurt?

[spuun]

Also maybe a timeout on the flush itself wouldn't hurt?

It wouldn't!