Re: Wrong mac in arp response in bonded interfaces

[Simon Chen <simonchennj@gmail.com>]

bond0 and the physical interfaces show up with the same mac in ifconfig...

The latest founding between me and my co-worker is that the MAC
address of the bonded interface switches between :44 and :45. Because
we're using an automated deployment tool, which unfortunately doesn't
reliably configure the udev rules correctly to persistent the two
NICs. So after a re-deploy, the MAC for the bonded NIC may switch. And
the stale MAC entry on the switch is then preventing return packet to
be delivered successfully.

oh, boy, this is just fun... will see if this is indeed the issue.

On Wed, Jan 18, 2012 at 2:47 PM, Jay Vosburgh <fubar@us.ibm.com> wrote:
> Simon Chen <simonchennj@gmail.com> wrote:
>
>>Hi all,
>>
>>Something really weird with interface bonding...
>>
>>I have eth0 and eth1, with MAC address xx:44 and xx:45. The bonded
>>interface chose to use xx:45 as its MAC.
>>
>>I configured an IP on the bonded interface, and try to ping the
>>default gw. The ARP from the server for the .1 is answered by the GW.
>>The server then sends out ICMP to the GW. The problem is the GW is not
>>responding to the ping.
>
>        How much real time is elapsing between the setting up of the
> bond, and this ping test?  What are the slaves set up as prior to the
> bond being established?  In particular, is one of them (the :44)
> assigned the IP address that the bond ends up using?
>
>>I then logged onto the GW (a switch) - apparently, the ARP table on
>>the GW shows that my server's IP is associated with xx:44 MAC address.
>>So, actually the GW is responding the ICMP, just to the wrong MAC
>>dest.
>>
>>Any idea how the xx:44 MAC somehow polluted the ARP table on my GW?
>>How can I make sure my server always sends out packets with xx:45 MAC
>>via the bonded interface?
>
>        My first suspicion is that a stale ARP entry on the switch is
> hanging around for the :44 MAC address from before the bond was
> established on the host.  If you clear the switch's ARP table, does the
> problem correct itself or happen again?
>
>        The other possibility that comes to mind is that you're using
> balance-alb mode, in which case I suspect what you're seeing is normal
> behavior.  The alb mode "assigns" peers to particular slaves of the bond
> by sending them tailored ARP messages bearing the MAC of one of the
> slaves, and each slave participates on the network under its own MAC
> address (I'm simplifying a bit here, but that's basically how it works).
>
>        -J
>
> ---
>        -Jay Vosburgh, IBM Linux Technology Center, fubar@us.ibm.com
>