[ovs-discuss] bonding switching strangely

Andreas Pflug pgadmin at pse-consulting.de
Thu Jul 4 13:03:17 UTC 2013


Hello,

I've setup some machines with debian 3.2.23 kernel and ovs 1.4.2 with
bonded interfaces, showing some strange behaviour:

# ovs-vsctl show
Bridge lan
        Port lan
            Interface lan
                type: internal
        Port "bond0"
            Interface "eth2"
            Interface "eth1"
        Port .....
   
# ovs-appctl bond/show bond0
bond_mode: active-backup
bond-hash-basis: 0
updelay: 0 ms
downdelay: 0 ms
lacp_negotiated: true

slave eth1: enabled
    active slave
    may_enable: true

slave eth2: enabled
    may_enable: true

#ifconfig lan
lan       Link encap:Ethernet  HWaddr x:x:x:x:x:5b
          inet addr:n.n.n.112  Bcast: .....


The machines have two dual-GBit interface cards, for example (on-board
82576 with igb driver):
card1:
  eth0  mac x:x:x:x:x:5a   -> not used
  eth1  mac x:x:x:x:x:5b   -> bonded to ovs bridge
card2:
  eth2  mac x:x:x:x:x:58   -> bonded to ovs bridge
  eth3  mac x:x:x:x:x:59   -> used elsewhere

Another one has one dual-NIC on-board one dual-NIC add-on (82576 and
82571EB with igb/e1000e drivers) with similar config. Note that in order
to improve redundancy, the bond is distributed between NIC interface
chips; if a chip fails both eth0/1 or eth2/3 might be affected at the
same time so the eth1/2 pair was chosen. Additionally, eth0 is also used
for an on-board IPMI device, which interferes with LACP so eth0 is
omitted from traffic.


Now, I infrequently see on the router
kernel: arp: n.n.n.112 moved from x:x:x:5b to x:x:x:x:x:5a on em1
kernel: arp: n.n.n.112 moved from x:x:x:5a to x:x:x:x:x:5b on em1

which means that the traffic temporarily moves away from eth1 to eth0,
which isn't a bond member at all! I do have mysteriuos network dropouts,
no traffic on some machines for some seconds, and a bond that's trying
to use a non-configured interface could well be the explanation. Can
anybody shed some light on this issue?

Regards,
Andreas





More information about the discuss mailing list