Bug #10203: TCPServer.new has strange behaviour when EADDRINUSE without specifying hostname

• Status: Closed
• Assignee: -
• Target version: -
• ruby -v: ruby 2.1.2p95
• Backport: 2.0.0: UNKNOWN, 2.1: UNKNOWN
• Tags: ext

[lucas (Lucas Nussbaum)]

after:
irb(main):003:0> TCPServer::new(10001)
=> #<TCPServer:fd 9>
I get irb listening to port 10001 using IPv4, but not IPv6:
tcp 0 0 0.0.0.0:10001 0.0.0.0:* LISTEN 1000 376068 24437/irb

a second creation also works, but only binds the IPv6 address:
irb(main):004:0> TCPServer::new(10001)
=> #<TCPServer:fd 10>
tcp6 0 0 :::10001 :::* LISTEN 1000 376098 24437/irb

=> I would have expected the first creation to bind to both IPv4 and IPv6, not just IPv4, and the second attempt to fail.

Trying once again, the creation fails with a strange exception:
irb(main):007:0> TCPServer::new(10001)
TypeError: no implicit conversion of nil into String
from (irb):7:in initialize' from (irb):7:in new'
from (irb):7
from /usr/bin/irb:11:in `'

Binding explicitely to 0.0.0.0 avoids this:
irb(main):005:0> TCPServer::new('0.0.0.0', 10002)
=> #<TCPServer:fd 11>
irb(main):006:0> TCPServer::new('0.0.0.0', 10002)
Errno::EADDRINUSE: Address already in use - bind(2) for "0.0.0.0" port 10002
from (irb):6:in initialize' from (irb):6:in new'
from (irb):6
from /usr/bin/irb:11:in `'

[naruse (Yui NARUSE)]

I know you was Debian porter, but show the actual ruby -v

[lucas (Lucas Nussbaum)]

that was from the Debian package indeed. I haven't rebuilt Ruby from source. The Debian package does not carry any Debian-specific patch at the moment. You cannot reproduce it?

[lucas (Lucas Nussbaum)]

Also, a colleague using Arch Linux checked (using the same Ruby version). He doesn't have IPv6 enabled.

And he also got the "TypeError: no implicit conversion of nil into String" exception instead of the Errno::EADDRINUSE one.

[akr (Akira Tanaka)]

• Status changed from Open to Feedback

It doesn't reproduce on my environment.

% lsb_release -idrc
Distributor ID:	Debian
Description:	Debian GNU/Linux testing (jessie)
Release:	testing
Codename:	jessie
% bin/irb          
irb(main):001:0> RUBY_VERSION
=> "2.2.0"
irb(main):002:0> require 'socket'
=> true
irb(main):003:0> TCPServer::new(10001)
=> #<TCPServer:fd 9>
irb(main):004:0> TCPServer::new(10001)
Errno::EADDRINUSE: Address already in use - bind(2) for nil port 10001
	from (irb):4:in `initialize'
	from (irb):4:in `new'
	from (irb):4
	from bin/irb:11:in `<main>'
irb(main):005:0> 

[lucas (Lucas Nussbaum)]

It might be fixed in Ruby 2.2... I don't have an easy way to test. Akira, do you have an easy way to test with Ruby 2.1?

If it is indeed fixed, it might be worth backporting to 2.1.

[normalperson (Eric Wong)]

[e-mail] wrote:

It might be fixed in Ruby 2.2... I don't have an easy way to test. Akira, do you have an easy way to test with Ruby 2.1?

If it is indeed fixed, it might be worth backporting to 2.1.

It may be fixed in trunk r44497.
Lucas: can you apply on your end and confirm? Thanks.

commit 5b0fb1aaddad03e111606e0d0eaf8ed9c9f7b0b6
Author: nobu nobu@b2dd03c8-39d4-4d8f-98ff-823fe69b080e
Date: Sat Jan 4 10:15:31 2014 +0000

 socket.c: format flags
 
 * ext/socket/socket.c (rsock_syserr_fail_host_port): use format flags,
   '+' to inspect, ' ' to quote unprintables.
 * ext/socket/socket.c (rsock_syserr_fail_path): ditto.
 * ext/socket/socket.c (rsock_syserr_fail_raddrinfo): ditto.
 
 git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@44497 b2dd03c8-39d4-4d8f-98ff-823fe69b080e

[akr (Akira Tanaka)]

Hm.

Ruby 2.1.3 causes the wrong error message:

% ./ruby -vrsocket -e 'p TCPServer::new(10001); p TCPServer::new(10001)'
ruby 2.1.3p242 (2014-09-19 revision 47630) [x86_64-linux]
#<TCPServer:fd 7>
-e:1:in `initialize': no implicit conversion of nil into String (TypeError)
	from -e:1:in `new'
	from -e:1:in `<main>'

However it doesn't succeed twice.

[lucas (Lucas Nussbaum)]

I built Ruby 2.2 from source, and still encountered the same problem as the original report, with the exception that the bogus error message seems fixed:
$ ./ruby -vrsocket -e 'p TCPServer::new(10001); system("netstat -ln|grep 10001"); p TCPServer::new(10001); system("netstat -ln|grep 10001"); p TCPServer::new(10001)'
ruby 2.2.0dev (2014-10-13 trunk 47902) [x86_64-linux]
#<TCPServer:fd 7>
tcp 0 0 0.0.0.0:10001 0.0.0.0:* LISTEN
#<TCPServer:fd 8>
tcp 0 0 0.0.0.0:10001 0.0.0.0:* LISTEN
tcp6 0 0 :::10001 :::* LISTEN
-e:1:in initialize': Address already in use - bind(2) for nil port 10001 (Errno::EADDRINUSE) from -e:1:in new'
from -e:1:in `'

I would have expected the first creation to bind to both IPv4 and IPv6, not just IPv4, and the second attempt to fail.

[lucas (Lucas Nussbaum)]

correct copy/paste:

$  ./ruby -vrsocket -e 'p TCPServer::new(10001); system("netstat -ln|grep 10001"); p TCPServer::new(10001); system("netstat -ln|grep 10001"); p TCPServer::new(10001)'
ruby 2.2.0dev (2014-10-13 trunk 47902) [x86_64-linux]
#<TCPServer:fd 7>
tcp        0      0 0.0.0.0:10001           0.0.0.0:*               LISTEN     
#<TCPServer:fd 8>
tcp        0      0 0.0.0.0:10001           0.0.0.0:*               LISTEN     
tcp6       0      0 :::10001                :::*                    LISTEN     
-e:1:in `initialize': Address already in use - bind(2) for nil port 10001 (Errno::EADDRINUSE)
	from -e:1:in `new'
	from -e:1:in `<main>'

[akr (Akira Tanaka)]

It seems net.ipv6.bindv6only affect the behavior.

% sudo sysctl net.ipv6.bindv6only=0
net.ipv6.bindv6only = 0
% ./ruby -vrsocket -e 'p TCPServer::new(10001); system("netstat -ln|grep 10001"); p TCPServer::new(10001); system("netstat -ln|grep 10001"); p TCPServer::new(10001)'
ruby 2.2.0dev (2014-10-13 trunk 47899) [x86_64-linux]
#<TCPServer:fd 7>
tcp        0      0 0.0.0.0:10001           0.0.0.0:*               LISTEN
-e:1:in `initialize': Address already in use - bind(2) for nil port 10001 (Errno::EADDRINUSE)
	from -e:1:in `new'
	from -e:1:in `<main>'

% sudo sysctl net.ipv6.bindv6only=1
net.ipv6.bindv6only = 1
% ./ruby -vrsocket -e 'p TCPServer::new(10001); system("netstat -ln|grep 10001"); p TCPServer::new(10001); system("netstat -ln|grep 10001"); p TCPServer::new(10001)'
ruby 2.2.0dev (2014-10-13 trunk 47899) [x86_64-linux]
#<TCPServer:fd 7>
tcp        0      0 0.0.0.0:10001           0.0.0.0:*               LISTEN
#<TCPServer:fd 8>
tcp        0      0 0.0.0.0:10001           0.0.0.0:*               LISTEN
tcp6       0      0 :::10001                :::*                    LISTEN
-e:1:in `initialize': Address already in use - bind(2) for nil port 10001 (Errno::EADDRINUSE)
	from -e:1:in `new'
	from -e:1:in `<main>'

This depends on OS configuration.

TCPServer.new doesn't hide this difference because some OS, such as OpenBSD, provides only the latter behavior.

I recommend Socket.tcp_server_sockets.
It returns two sockets for IPv4 and IPv6, regardless of net.ipv6.bindv6only.

% sudo sysctl net.ipv6.bindv6only=0
net.ipv6.bindv6only = 0
% ./ruby -rsocket -e 'p Socket.tcp_server_sockets(10001)'
[#<Socket:fd 7>, #<Socket:fd 8>]
% sudo sysctl net.ipv6.bindv6only=1
net.ipv6.bindv6only = 1
% ./ruby -rsocket -e 'p Socket.tcp_server_sockets(10001)'
[#<Socket:fd 7>, #<Socket:fd 8>]

[lucas (Lucas Nussbaum)]

Indeed, bindv6only=1 seems to change this. However, bindv6only=1 is kind-of the default on all systems (except some Linux distros).

My goal was to implement the following:
try to bind port 10001 ; if already taken, bind port 10002
That's not possible with the current behaviour.

I don't think it's an OS behaviour. I straced the ruby process, and saw:

first TCPServer call:
[pid 16280] bind(7, {sa_family=AF_INET, sin_port=htons(10001), sin_addr=inet_addr("0.0.0.0")}, 16) = 0
second TCPServer call:
[pid 16280] bind(8, {sa_family=AF_INET, sin_port=htons(10001), sin_addr=inet_addr("0.0.0.0")}, 16) = -1 EADDRINUSE (Address already in use)
[pid 16280] bind(8, {sa_family=AF_INET6, sin6_port=htons(10001), inet_pton(AF_INET6, "::", &sin6_addr), sin6_flowinfo=0, sin6_scope_id=0}, 28) = 0
third TCPServer call:
[pid 16280] bind(9, {sa_family=AF_INET, sin_port=htons(10001), sin_addr=inet_addr("0.0.0.0")}, 16) = -1 EADDRINUSE (Address already in use)
[pid 16280] bind(9, {sa_family=AF_INET6, sin6_port=htons(10001), inet_pton(AF_INET6, "::", &sin6_addr), sin6_flowinfo=0, sin6_scope_id=0}, 28) = -1 EADDRINUSE (Address already in use)

So the second bind() on AF_INET port 10001 fails correctly. It seems that Ruby implements a fallback mechanism that makes it bind the corresponding AF_INET6 port. I think that's wrong. But maybe it's just a documentation issue.

[akr (Akira Tanaka)]

Ruby uses getaddrinfo() to obtain addresses to bind.

% ./ruby -rsocket -e 'p Addrinfo.getaddrinfo(nil, 10001, nil, :STREAM, nil, Socket::AI_PASSIVE)'
[#<Addrinfo: 0.0.0.0:10001 TCP>, #<Addrinfo: [::]:10001 TCP>]

TCPServer.new returns the first socket succeed to bind.

If net.ipv6.bindv6only=1, the first TCPServer.new returns IPv4 socket and the second returns IPv6 socket.

If net.ipv6.bindv6only=0, first TCPServer.new returns IPv4 socket and second fails because
IPv6 socket for 10001 needs IPv4 10001 but it is already used, I guess.

The biggest problem here is TCPServer.new returns only one socket.
I think TCP server should listen all addresses obtained by getaddrinfo() but it is impossible with TCPServer.new().

So, my suggestion again:
I recommend Socket.tcp_server_sockets.

[lucas (Lucas Nussbaum)]

I was not interested in adding IPv6 support to my application, only in binding a port with a fallback mechanism. I worked around the problem with bind('0.0.0.0', 10001).

But I still think that it is wrong that TCPServer.new implements this fallback mechanism. It would be less surprising if it just tried to bind the first address, and fail if unsuccessful. I can't think of a scenario where this is useful.

[akr (Akira Tanaka)]

I'm not sure that ignoring getaddrinfo()'s non-first entries is a good behavior.

[jeremyevans0 (Jeremy Evans)]

• Status changed from Feedback to Closed

I think this is not a bug and can be closed. TCPServer.new using first getaddrinfo address that works is reasonable behavior if no address is specified. If you require a specific address, you should specify it. As @akr (Akira Tanaka) mentioned, Socket.tcp_server_sockets is a better way to handle the situation if you want to listen on both IPv4 and IPv6 addresses.