SO_REUSEADDR should not also mean SO_REUSEPORT

[Sean McNeil]

On Tue, 2006-02-21 at 10:11 -0500, Kurt Miller wrote:
> On Saturday 18 February 2006 3:45 pm, Arne H. Juul wrote:
> > On Sat, 18 Feb 2006, Nate Williams wrote:
> > >> Ok, thanks. I got that impression from reading some posts I found
> > >> while googling. There was one in particular for NetBSD that
> > >> discussed it in detail. Check out the Apr 2 portion of this
> > >> http://www.tinyurl.com/b46gq by Jan Schaumann. Also this
> > >> one http://tinyurl.com/9sa6a. From these posts it appears
> > >> that SO_REUSEPORT is needed in some cases to be compatible
> > >> with linux.
> > >
> > >> From the early days....
> > >
> > >  - In the Multicast constructor, the low level routine sets the
> > >    SO_REUSEADDR option by using JSO_REUSEADDR which corresponds to a call
> > >    to setsockopt(..SO_REUSEADDR).  To make multicast sockets work in *all*
> > >    cases on FreeBSD, we should also set SO_REUSEPORT, else in many cases
> > >    the multicast bind will fail.
> > 
> > I won't claim to know what's the best behaviour with multicast, but the 
> > problem is that SO_REUSEPORT is always used when SO_REUSEADDR was 
> > requested, meaning that:
> > 
> > > SO_REUSEPORT allows completely duplicate bindings by multiple
> > > processes if they all set SO_REUSEPORT before binding the port.
> > 
> > so you can have two very different java servers listening on the same 
> > port, for example.  Or the same java server started twice won't notice any 
> > problem because the second instance will bind its server port fine, while 
> > on all other OSes this would give a sensible error message.  And so on. 
> > This is bad.
> > 
> > The reason I found this problem in the first place was from a Java program 
> > that worked well on Linux, not at all on FreeBSD, and after much tracing 
> > we deduced that something was enabling SO_REUSEPORT on FreeBSD, after 
> > which finding the bad code was a simple matter of "grep", only leaving the 
> > question of why it was there in the first place.
> > 
> > If anybody figures out what's best practice for supporting multicast 
> > applications, ask the BSD kernel people to change the kernel behaviour to 
> > match best practice, make it possible to control SO_REUSEPORT from the 
> > MulticastSocket class, or find some other solution that doesn't make 
> > *other* types of java application suffer.
> 
> Thanks for the explanation and also to Nate for the Multicast
> history. I've looked into this a bit more over the weekend and
> found that the network stack promotes SO_REUSEADDR to include
> SO_REUSEPORT for multicast addresses, so I believe that case is
> covered already. I ran the network jck's on the 1.5 jvm with your
> patch and found that SO_REUSEPORT is still needed to pass the jck's
> but for datagram sockets only.
> 
> Could you try this patch and test it with the program you referred
> to above?
> 
> --- ../../j2se/src/solaris/native/java/net/net_util_md.c.orig	Tue Feb 21 09:56:11 2006
> +++ ../../j2se/src/solaris/native/java/net/net_util_md.c	Tue Feb 21 10:06:31 2006
> @@ -1022,11 +1022,20 @@
>      }
>  
>      /*
> -     * If SO_REUSEADDR option requested, unconditionally set SO_REUSEPORT.
> +     * If SO_REUSEADDR option requested for SOCK_DGRAM, set SO_REUSEPORT also.

All UDP sockets?  Why not just test the address to see if it is a
multicast address?  Shouldn't SO_REUSEPORT be set for TCP in that case
as well?

Sean