[Lxc-users] Bug discussion: implementing high virtual device MAC addresses

Mon Oct 24 22:49:29 UTC 2011

     Serge,
     Thank you for looking at this.

Serge> /However, I actually don't think it should happen the way you 
describe./

     I believe you have mis-read my description.  I think we are 
actually in agreement with what is happening.

     You said:

Serge> /So the mac address of the veth endpoint in the container should 
not matter./

     I think that is the same thing that I said:

Derek> [The problem MAC address] is NOT the mac address specified in 
lxc.conf, like this:

lxc.network.hwaddr = fe:16:3e:fd:5a:5b

	That MAC address has nothing to do with the bug; the host's bridge
device (br0) will never assume a configured LXC MAC address as its own.

     Also, you said:

Serge> /The other endpoint, the veth which stays in the host's network 
namespace, that is the one which gets placed on the bridge./

     I agree, that is the address which causes the ~4 network second 
freeze.  As I said in my original description:

Derek>> ...the MAC address in question is the one of the virtual 
vethXXXX device, as shown with "ifconfig" on the host:

veth0IEDlk Link encap:Ethernet  HWaddr 4e:34:7c:dc:92:e8
[...snip...]

     So, are we in agreement that the problem address is NOT the one in 
the LXC .conf file (as specified by the user), but instead is the 
"random" address of the veth device on the host?

Serge> /Hmm, I haven't seen this happen at all./

     I have seen it on Ubuntu 10.04, and there was an independent 
description of the same symptom (and a different but very similar 
work-around) filed in SourceForge here:

http://sourceforge.net/tracker/index.php?func=detail&aid=3411497&group_id=163076&atid=826303

     (That's SF bug ID# 3411497.)

     As described in the libvirt bugfix for this issue (linked below), 
the reason some people see it and some people don't is that it only 
happens when the veth MAC address is lower than that of the physical 
eth0 device's MAC address.  (That is how the Linux kernel handles it, by 
design.  I don't know why.)

     Since the MAC address is randomly chosen, it is a random symptom 
that will vary from one NIC to another.  Those who happen to have a high 
MAC address for eth0 will see it more frequently (but still randomly.)  
This is a major impact on production symptoms, where a ~4 second network 
freeze could trigger admin alerts and/or failover scripts.  (Note the 
exact duration of the network freeze also depends on your switches and 
routers, and how they handle ARP caching.)

Thank You,
Derek Simkowiak

On 10/24/2011 11:40 AM, Serge E. Hallyn wrote:
> Quoting Derek Simkowiak (derek at simkowiak.net):
>>       Hello,
>>       Just following up re: this bug.  I think it's a pretty serious issue.
>>
>>       I am looking to work on this, but I am seeking some feedback and
>> direction from one of the core LXC devs.
>>
>> - Do you agree with my analysis?
>> - Has anyone else worked on this already?
> Hmm, I haven't seen this happen at all.  That doesn't mean it's not
> possible.
>
> However, I actually don't think it should happen the way you describe.
> Note that the veth passed in to the container is *not* assigned to the
> bridge.  The other endpoint, the veth which stays in the host's network
> namespace, that is the one which gets placed on the bridge.  So the
> mac address of the veth endpoint in the container should not matter.
>
> (Disclaimer: my being wrong is a not-infrequent event)
>
> -serge
>
>> etc.
>>
>>
>> Thanks,
>> Derek
>>
>> On 10/18/2011 04:31 PM, Derek Simkowiak wrote:
>>>        There is a behavior in the Linux kernel which can cause a bridge
>>> device to change MAC address, thus causing a network blackout of several
>>> seconds (while everybody ARPs the new MAC address flushes the old one).
>>> This happens when bridging an enslaved interface, like we do with LXC.
>>>
>>>        The symptom is that the LXC host will black out for several seconds
>>> when starting or stopping an LXC container.  Your SSH terminal on the
>>> host will freeze and become unresponsive.  (It is a random symptom,
>>> because the blackout only happens if the randomly-assigned MAC address
>>> of the virtual device is lower than that of the physical eth0 device).
>>>
>>>        This behavior was first observed by the libvirt folks when creating
>>> virtual machines.  You can read more details about it (and how they
>>> fixed it) here:
>>>
>>> https://www.redhat.com/archives/libvir-list/2010-July/msg00450.html
>>> https://bugs.launchpad.net/ubuntu/+source/libvirt/+bug/584048
>>>
>>>        I have observed the symptom under LXC, and the workaround for it
>>> has been independently confirmed for LXC in this bug report (ID: 3411497):
>>>
>>> http://sourceforge.net/tracker/index.php?func=detail&aid=3411497&group_id=163076&atid=826303
>>>
>>>
>>>        The workaround for the bug is to give the virtual device a high MAC
>>> address, thus discouraging the bridge device from adapting its MAC
>>> address as its own.
>>>
>>>        I have mentioned this bug on the list before, however, I was
>>> confused about which MAC address was causing the problem.  This is NOT
>>> the mac address specified in lxc.conf, like this:
>>>
>>> lxc.network.hwaddr = fe:16:3e:fd:5a:5b
>>>
>>>        That MAC address has nothing to do with the bug; the host's bridge
>>> device (br0) will never assume a configured LXC MAC address as its own.
>>> Instead, the MAC address in question is the one of the virtual vethXXXX
>>> device, as shown with "ifconfig" on the host:
>>>
>>> veth0IEDlk Link encap:Ethernet  HWaddr 4e:34:7c:dc:92:e8
>>> [...snip...]
>>>
>>>        That HWaddr should be given a high prefix to avoid the network
>>> blackouts, just like they've done for libvirt.  That does not exist in
>>> any config file anywhere; it must be fixed in the LXC source code.
>>>
>>>        I looked in network.c for the LXC source code and I think the fix
>>> should go in lxc_bridge_attach() near line 991.  The fix would put a
>>> manually-generated MAC address -- one with a high prefix -- into
>>> ifr.ifr_hwaddr.sa_data and thus replace the random one assigned by the
>>> kernel.
>>>
>>>        However, I'm new to the LXC source and would like some input and
>>> analysis from a more seasoned contributor.  I would be happy to test and
>>> maybe even contribute a patch, but I'd like some feedback first.
>>>
>>>
>>> Thank You,
>>> Derek Simkowiak
>>>
>>>
>>> ------------------------------------------------------------------------------
>>> All the data continuously generated in your IT infrastructure contains a
>>> definitive record of customers, application performance, security
>>> threats, fraudulent activity and more. Splunk takes this data and makes
>>> sense of it. Business sense. IT sense. Common sense.
>>> http://p.sf.net/sfu/splunk-d2d-oct
>>> _______________________________________________
>>> Lxc-users mailing list
>>> Lxc-users at lists.sourceforge.net
>>> https://lists.sourceforge.net/lists/listinfo/lxc-users
>>
>> ------------------------------------------------------------------------------
>> The demand for IT networking professionals continues to grow, and the
>> demand for specialized networking skills is growing even more rapidly.
>> Take a complimentary Learning at Cisco Self-Assessment and learn
>> about Cisco certifications, training, and career opportunities.
>> http://p.sf.net/sfu/cisco-dev2dev
>> _______________________________________________
>> Lxc-users mailing list
>> Lxc-users at lists.sourceforge.net
>> https://lists.sourceforge.net/lists/listinfo/lxc-users

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.linuxcontainers.org/pipermail/lxc-users/attachments/20111024/9511e07b/attachment.html>