[Cerowrt-devel] mdns reflector issues on ipv6/babel routing through nat.

[Dave Taht]

On Sun, Mar 18, 2012 at 2:27 PM, Jim Gettys <jg at freedesktop.org> wrote:
> Oh, cool.  Broadcast storms are soooo much fun...

I ended up having to disable avahi entirely.

the NEAT thing about this, was that although my network was basically
saturated, it still more or less worked and I hardly noticed I had a
problem, due to having sfq everywhere....

>                - Jim
>
>
> On 03/18/2012 05:24 PM, Dave Taht wrote:
>> On Sun, Mar 18, 2012 at 2:22 PM, Dave Taht <dave.taht at gmail.com> wrote:
>>> Once you get to a few routers, a few deep, (3 in series in this case,
>>> 9 overall), the avahi mdns proxy starts to malfunction over ipv6, and
>>> I ended up with a rather nasty broadcast storm.
>>>
>>> So I had to disable the ipv6 multicast of mdns in order to get my
>>> network back in this (excessively) complex network.
>>>
>>> use-ipv6=no in the /etc/avahi/avahi-daemon file
>>>
>>> Seems to work fine, two deep. Curiously, I did not observe a similar
>>> storm for ipv4...
>>>
>>> Now this is across like 5 different versions of cerowrt, but it would
>>> not surprise me that this is a  generic problem with avahi on ipv6,
>>> and/or a symptom of the brain-damaged-ness of mdns in the first place.
>>>
>>> use-ipv6=no
>>>
>>> I note that when you connect cero boxes together in a babel mesh
>>> configuration, site-local multicast is not a problem, because it
>>> doesn't work in the first place (by design). This can be construed as
>>> an advantage (no broadcast storm), or disadvantage (mdns and
>>> site-local multicast doesn't work across meshed links)
>> Actually  I was wrong. I'm STILL observing a broadcast storm, AND it
>> is taking place across the meshed links too.... aggggh.....
>>
>>> Incidentally, I don't know if anyone would purposely inflict a network
>>> this complex on themselves:
>>>
>>> http://pastebin.com/LzeeiCXg
>>>
>>> but it does illustrate that a complex, automagically routed, fault
>>> tolerant ipv4 and ipv6 network IS feasible, so long as all internal
>>> addresses are unique.
>>>
>>> The biggest problem I run into is that 'failover-capable,
>>> fault-tolerant routing' introduces major headaches with firewall
>>> rules.
>>>
>>> Another thing the above paste illustrates that you can mix and match
>>> ipv4 nat with ipv6 fully meshed routing.
>>>
>>> The box I took that trace off has babel enabled on all interfaces, and
>>> has the following rule at the top of it's babeld.conf file
>>>
>>> out if ge00 ip 0.0.0.0/0 deny
>>>
>>> (as do multiple other boxes in the lab on the external network)
>>>
>>> this prohibits announcing ipv4 routes across the natted ge00
>>> interface, but allows ipv6. In the caseof that paste, this particular
>>> router has NO internal wired connections at all, it just meshes
>>> internally for ipv4, and because ge00 is a higher quality (ethernet)
>>> interface, babel chooses it for the default for ipv6 for most routes.
>>>
>>>
>>>
>>> --
>>> Dave Täht
>>> SKYPE: davetaht
>>> US Tel: 1-239-829-5608
>>> http://www.bufferbloat.net
>>
>>
>



-- 
Dave Täht
SKYPE: davetaht
US Tel: 1-239-829-5608
http://www.bufferbloat.net