Bug 8349

Reply-To: akpm@linux-foundation.org

On Tue, 17 Apr 2007 18:45:12 -0700 bugme-daemon@bugzilla.kernel.org wrote:

> http://bugzilla.kernel.org/show_bug.cgi?id=8349
> 
>            Summary: Manually set IPv6 default route ignored
>     Kernel Version: 2.6.20.7
>             Status: NEW
>           Severity: normal
>              Owner: yoshfuji@linux-ipv6.org
>          Submitter: linkfanel@yahoo.fr
> 
> 
> Most recent kernel where this bug did *NOT* occur: 2.6.19.5
> Distribution: Debian sid
> Hardware Environment: x86 (Pentium 4/MMX)
> Software Environment: iproute2-ss060323, ifupdown 0.6.8
> 
> Problem Description:
> Default routes added to the IPv6 routing table, which are not coming from
> autoconfiguration, seem to be ignored. Subsequently, IPv6 routing is disrupted,
> and for instance calls to connect() fail with an ENETUNREACH error.
> 
> Steps to reproduce:
> 0. Find an IPv6-enabled box on an IPv6-free network. In the absence of any other
> configuration, the routing table contains the single line:
> fe80::/64 dev eth0 ...
> 1. Add a default route:
> ip -6 ro add default via fe80::1 dev eth0
> 2. Try to do IPv6, for instance:
> # ping6 2001:4978:1:110:0:ac:ce55:1b1e
> connect: Network is unreachable
> (returns instantly, no packet is sent)
> 
> This bug is present in the last stable 2.6.20.7 kernel, and in 2.6.21-rc7. I
> actually came across this bug on a server on a network where IPv6 is routed and
> advertised, but with autoconfiguration disabled on this server. It *seems* to me
> that if the interface is brought up with accept_ra enabled, default routes,
> including the one got from RA and ones input by hand later will work, whereas if
> it is brought up with accept_ra disabled (or no RA is received), the problem
> will occur. Note that routes with a non-null prefix length, such as 3 or 1, work
> fine.

Verified for 2.6.21.3, a box connected to a network segment with two IPv6
routers (both with no router advertisements) which is learning a default route
with Quagga/RIPngd fails to use the default route installed into the kernel by
quagga.

svr02:~# ip -6 route
[...]
default via fe80::20c:86ff:fe9a:3819 dev vlan15  proto zebra  metric 2  expires
21334056sec mtu 1500 advmss 1440 hoplimit 4294967295
svr02:~# ping6 2001:a60:f001:1:218:f3ff:fe66:c777
connect: Network is unreachable
svr02:~# ip -6 route add 2000::/3 via fe80::20c:86ff:fe9a:3819 dev vlan15
svr02:~# ping6 2001:a60:f001:1:218:f3ff:fe66:c777
PING 2001:a60:f001:1:218:f3ff:fe66:c777(2001:a60:f001:1:218:f3ff:fe66:c777) 56
data bytes
64 bytes from 2001:a60:f001:1:218:f3ff:fe66:c777: icmp_seq=1 ttl=59 time=7.67 ms

According to various reports on IRC and other channels, 2.6.20.1 and 2.6.20.3
work fine, according to the original bugreport 2.6.20.7 is broken.

There is a rather large patch to net/ipv6/route.c for 2.6.20.5, which might be
causing this regression.

http://www.kernel.org/diff/diffview.cgi?file=%2Fpub%2Flinux%2Fkernel%2Fv2.6%2Fincr%2Fpatch-2.6.20.4-5.bz2;z=53

I can't really tell whether the problem is caused by this patch, but it looks
like a good candidate to me.

Pierre, since that bug affects several people and it is a regression (introduced
 in -stable and apparently still in the wild), could you please raise the severity?

Thanks,
Bernhard

No, it is not appropriate to raise the severity just because it is a regression
in -stable.

This bug exists in 2.6.21 too.

The patch[1] which causes this to break is being debugged in this thread:
http://lkml.org/lkml/2007/5/28/299

1. commit f11e6659ce9058928d73ff440f9b40a818d628ab [IPV6]: Fix routing
round-robin locking.

The commit

commit b407fdbabb6c630660e1e9c9c2d212a03b746775
Author: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Date:   Wed Jun 6 22:42:58 2007 -0700

    [PATCH] IPV6 ROUTE: No longer handle ::/0 specially.
    
    We do not need to handle ::/0 routes specially any longer.
    This should fix BUG #8349.

did fix it for me. I suggest this bug be closed.