Crystal 0.20 Error while spawning 2 HTTP::Server on the same port - SO_REUSEADDRESS

[adrien-thierry]

I'm trying to make a multithreaded/forked http server, but I can't set reuse_address, and it should be true by default.

Code :

require "http/server"
spawn do
  server = HTTP::Server.new(8080) do |context|
    context.response.content_type = "text/plain"
    context.response.print "Hello world, got #{context.request.path}!"
  end
  puts "Listening on http://127.0.0.1:8080"
  server.listen
end

spawn do
  server = HTTP::Server.new(8080) do |context|
    context.response.content_type = "text/plain"
    context.response.print "Hello world, got #{context.request.path}!"
  end
  puts "Listening on http://127.0.0.1:8080"
  server.listen
end

sleep

Error :

Error binding TCP server at 127.0.0.1:8080: Address already in use (Errno)
0x470917: *CallStack::unwind:Array(Pointer(Void)) at ??
0x4708aa: *CallStack#initialize:Array(Pointer(Void)) at ??
0x47087a: *CallStack::new:CallStack at ??
0x467bde: *raise<Errno>:NoReturn at ??
0x517e34: *TCPServer#initialize<String, Int32, Int32>:Nil at ??
0x5178c1: *TCPServer#initialize<String, Int32>:Nil at ??
0x517874: *TCPServer::new<String, Int32>:TCPServer at ??
0x50fbfc: *HTTP::Server#bind:TCPServer at ??
0x50fb2a: *HTTP::Server#listen:Nil at ??
0x46dcf8: ~procProc(Nil) at ??
0x47fa10: *Fiber#run:(IO::FileDescriptor | Nil) at ??
0x46b0b6: ~proc2Proc(Fiber, (IO::FileDescriptor | Nil)) at ??
0x0: ??? at ??

in crystal/src/socket/tcp_server.cr :

Line 20 : self.reuse_address = true

[ysbaddaden]

There is a misunderstanding in how SO_REUSEADDR works, and binding to the exact same IP:PORT is invalid. This (incredible) stackoverflow answer explains why: http://stackoverflow.com/questions/14388706/socket-options-so-reuseaddr-and-so-reuseport-how-do-they-differ-do-they-mean-t#14388707

[adrien-thierry]

Thx, I read it

[Sija]

@ysbaddaden is it possible to use SO_REUSEPORT with HTTP::Server ATM?

[adrien-thierry]

@Sija @ysbaddaden I've compiled crystal with SO_REUSEPORT on Linux x64, works great, but not tested on other platforms

[ysbaddaden]

@Sija there is no limitation to use it, as long as you can set the option, which I'm unsure with the current Crystal API.

BTW, if I recall correctly, we do set SO_REUSEADDR to be able to immediately rebind a socket that was just unbound, but still lingering (e.g. restarting a server), which would prevent binding for at least one minute. We don't set SO_REUSEPORT because we don't expect this behavior by default.

[luislavena]

@ysbaddaden @bcardiff perhaps we can introduce Socket#reuse_port= and Socket#reuse_port? and later expose that something like shared: true on the HTTP::Server interface? (just a thought)

[adrien-thierry]

If someone want to test on other platform :

self.reuse_port = true under self.reuse_address = true in class TCPServer in txp_server.cr

SO_REUSEPORT = 15 under SO_REUSEADDR = 2 in LibC binding in /sys/socket.cr

And this :

def reuse_port?
  getsockopt_bool LibC::SO_REUSEPORT
end

def reuse_port=(val : Bool)
  setsockopt_bool LibC::SO_REUSEPORT, val
end

in class Socket in socket.cr

I used this to benchmark Crystal with multithreading NodeJS vs Crystal vs Golang