Re: TUN/TAP/UML proxy arp and IPV6 don't like one another.

From: Willy Tarreau
Date: Wed Nov 15 2006 - 00:34:24 EST


Hi,

On Tue, Nov 14, 2006 at 09:30:26AM -0600, greg@xxxxxxxxxxxx wrote:
> On Nov 8, 1:16pm, Jeff Dike wrote:
> } Subject: Re: TUN/TAP/UML proxy arp and IPV6 don't like one another.
>
> Good morning Jeff, thanks for the reply.
>
> I've added Willy as a CC to this note to bring him in on the issue.
>
> Willy I'm able to demonstrate a problem with ARP in the 2.4.33.3 which
> seems to be related to the presence of IPV6 in the kernel. Details
> below.

unfortunately, I don't see many details, but only some parts of mails
already exchanged. From what I've read, you seem to have a problem
related proxy ARP on IPv4 when enabling IPv6. Could you please :

1) give me a short description of what problem you observe and how
you can observe it.

2) tell me if you remember of a previous 2.4 which did not exhibit
the problem. If you don't know, please check 2.4.33.2 since which
a patch affecting ARP was merged.

3) if you have already identified a test case which show the problem
without installing a full UML environment, can you reproduce the
problem without the UML patches ? Not that I point the finger to
the UML patches at all, but this point is important to help
reproduce in the most general case, or identify parts of the kernel
touched by the UML patches which have changed since last working
version.

> > On Fri, Nov 03, 2006 at 12:08:01PM -0600, greg@xxxxxxxxxxxx wrote:
> > > Environment:
> > > 2.4.33.3 host kernel with SKAS3 patch.
> > > Stock 2.6.18 UML guest kernel.
> > >
> > > Host kernel has IPV6 enabled but there are no IPV6 routers
> > > active on the subnet. The active eth0 host interface only has
> > > a link-local address assigned to it.
> > >
> > > The host kernel was recompiled with the only change being to drop IPV6
> > > from the kernel. Identical guest OS implementation boots and properly
> > > pings the same target host on the network.
>
> > So, you're saying that IPV6, even though it is not being used, is
> > somehow interfering with normal IPV4 networking?
>
> That seems to be the case. Most specifically the willingness of the
> kernel to proxy-ARP for an IPV4 address assigned to the end of a
> TUN/TAP interface when IPV6 is enabled in the host kernel.
>
> With everything else held constant removing IPV6 from the host kernel
> makes the problem go away.
>
> > > I know 2.4.33.3 is a bit dated so this might be better to bounce off
> > > Willy.
> > >
> > > Thoughts, further suggestions?
>
> > Maybe upgrade and hope the problem goes away. It sounds like this
> > should be reproducable without UML, so if you can reproduce this on
> > a newer kernel with a smaller test case, that might help get it
> > fixed.
>
> I'm including Willy on the mail since if its a regression in the 2.4.x
> its probably worth chasing down. 2.6 is nice but we have literally a
> ton of 2.4 running for stability reasons.

>From experience, you should avoid publically justifying the use of 2.4 by
"for stability reasons" when looking for help, as it seems to make some
developers nervous ; saying "for various reasons" hurts less people :-)

> Willy any thoughts on a test harness for this problem beyond UML? I
> have the UML environment we use all bottled up into an RPM with a
> reasonably sophisticated startup script. Invoking a UML environment
> to validate the bug is basically an RPM install followed by a single
> command. I can easily drop the RPM on an FTP server.

Once I understand your problem, we might have to try such a method,
but let's see if we can reproduce in an easier manner.

Regards,
Willy

-
To unsubscribe from this list: send the line "unsubscribe linux-net" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html