Re: [PATCH 1/2] bridge: leave carrier on for empty bridge

[Ang Way Chuang <wcang@sfc.wide.ad.jp>]

On 05/09/11 02:12, Krzysztof Olędzki wrote:
> On 2011-09-04 18:36, Stephen Hemminger wrote:
>> On Sun, 04 Sep 2011 09:35:10 +0200
>> Nicolas de Pesloüan<nicolas.2p.debian@gmail.com>  wrote:
>>
>>> Le 04/09/2011 06:14, Stephen Hemminger a écrit :
>>>
>>>>> Instead of asserting carrier when the bridge have no port, can't we assert carrier when the three
>>>>> following condition are true at the same time :
>>>>>
>>>>> - The bridge have no port.
>>>>> - At least one IP address is setup on the bridge.
>>>>> - The two above conditions are true for more than a configurable amount of seconds, with a default
>>>>> of 10, for example.
>>>>>
>>>>> This would only delay carrier on for a few seconds for the regression and keep the current behavior
>>>>> (carrier off until at least 1 port is on) for DHCP.
>>>>
>>>> This fails on two counts:
>>>> 1. Bridge's often run without IP addresses!
>>>> 2. DHCP won't try and send out request until carrier is true.
>>>
>>> Sorry, I missed to say that we should of course also assert carrier on if one port has carrier on.
>>>
>>> And rethinking about it, the delay is probably useless :
>>>
>>> bridge_carrier_on = at_least_one_port_has_carrier_on | (bridge_has_no_port&  bridge_has_at_least_one_ip)
>>>
>>> That way :
>>> - for those using bridge without any port, manually setting the IP will assert carrier on. (By the
>>> way, why don't they use a dummy device instead?)
>>>
>>> - for those using bridge with ports:
>>> -- Using any kind of autoconfig will work as expected. Carrier will only be asserted at the time
>>> first port get carrier.
>>> -- Using static IP confifiguration, carrier will possibly be erroneously reported as on during the
>>> small time gap between IP address configuration and first port is added to the bridge. This time gap
>>> may be removed by simply configuring the IP after the first port is added. This is probably already
>>> true for most distribs. And anyway, this time gap is probably not a problem.
>>> -- Carrier will also be erroneously reported as on after removing the last port, if the bridge still
>>> has an IP. (But we can arrange for this not to happen).
>>>
>>> And in order to ensure user really understand why carrier is on of off, we can simply issue an INFO
>>> message for the non-natural case (bridge_has_no_port&  bridga_has_at_least_one_ip).
>>>
>>> I consider all this reasonable.
>>>
>>>     Nicolas.
>>
>> Any bridge behaviour based on IP address configuration is a
>> layering violation and won't work.  The problem is related to dynamic issues
>> with IPv6 and DHCP and needs to be addressed at that level.
>
> Maybe we can simply add a switch controlling if a bridge with no attached ports has carrier off (default) or on.
>
> For example:
>  echo {0|1} > /sys/devices/virtual/net/brX/bridge/orphaned_carrier
>
>  brctl orphaned_carrier brX {on|off}
>
> Best regards,
>
>                 Krzysztof Olędzki
My colleague investigated the problem. In the case of libvirt, there are 2 issues at hand:
- dnsmasq cannot bind to port 53 for IPv6 address, therefore fails.
- radvd fails to start because the interface is treated as being down.

This blog entry from Daniel Berrange explains the configuration used by libvirt:
http://berrange.com/posts/2011/06/16/providing-ipv6-connectivity-to-virtual-guests-with-libvirt-and-kvm/
It's quite possible that other users may have the same problem in the future if they plan to use radvd or
dnsmasq on a bridge interface. I'm not sure if there are other applications that depend on the carrier on
a bridge interface to function properly. Come to think of it, unless the user is aware of the change
in carrier on the bridge interface, it is very hard for them to diagnose the problem and apply the right
solution (toggling the carrier). Ideally, there is no reason why dnsmasq and radvd should be started if
there is no port attached to a bridge interface. But, we cannot assume that's how it is done.

I think that it is prudent that the whole commit should be reverted because the problem may not be limited
to libvirt use case.
>
> -- 
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>