your mail

[Maxim Dounin]

Hello!

On Wed, Jun 01, 2011 at 10:33:42AM -0400, Les Aker wrote:

> I've found what appears to be a bug in the handling of listen 
> directives for ipv4 and ipv6.
> 
> When running two vhosts, one of which listens on [::], the other 
> on localhost, nginx will fail with this error:
> 2011/06/01 10:12:35 [emerg] 9750#0: bind() to 127.0.0.1:80 failed (98: Address already in use)
> My understanding was that this error stems from [::] being 
> inclusive of 127.0.0.1, causing nginx to attempt to bind it 
> twice.
> However, replacing the `listen localhost;` with `listen [::1];` 
> removes the error, despite the fact that [::1] is also included 
> within [::].  -- Les Aker

Linux, I suppose?  There are two problems here:

1. Linux doesn't allow multiple bind()'s on the same port, even 
with setsockopt(SO_REUSEADDR).

2. nginx doesn't expect [::] to conflict with / include 127.0.0.1, 
as these two belong to different address families.  But in fact 
depending on v6only system (or socket) setting [::] may either 
include 127.0.0.1 or not.

With [::] and [::1] nginx expects conflict to happen and 
workarounds it for Linux by using single bind() on [::] and 
getsockaddr() call on each new connection.  With [::] and 
127.0.0.1 it doesn't (see (2)) and that's why you see conflict.

Simple solution is to explicitly listen on ipv4 and ipv6 (with 
v6only option) sockets, like this:

    listen *:80;
    listen [::]:80 ipv6only=on;

(actually, this is how it works on other operating systems 
with v6only being on by default, and either on Linux with 
appropriate sysctl being set)

Maxim Dounin