Dear LVS users / netdev readers, today I've got a pretty peculiar problem.
I've been running 3.14.48 (and some earlier 3.14 kernels) for a long time now in an LVS / keepalived driven loadbalancing cluster. See below for more detail on the setup. Today I started to upgrade to the current 3.14.70 kernel. At first glance everything seems fine, I can failover _to_ the box with the new kernel, and traffic is flowing fine. However, when I then switch BACK to a different box, the 3.14.70 kernel crashes. I've got an incomplete console dump (IPMI avi capture, single frame...) you can see here: https://plus.google.com/u/0/photos/photo/114613285248943487324/6288270822391242162 I usually manually failover by (having some automation) fiddle with keepalived.conf VRRP priority settings, then restart the daemon. The issue / reboot only manifests when I RESTART the ACTIVE keepalived. It always happens, then, with 3.14.70. That never happened before. On the other hand just reloading keepalived, with the prio-modified config, works fine! As I don't remember why I restarted instead of reloading, I can for now change my automation easily - but the issue is weird anyway. More info on the setup: 1) kernel is vanilla 3.14.70 (and was vanilla 3.14.48 without the issue), with a single (self written) patch to bonding applied (see http://permalink.gmane.org/gmane.linux.network/316758). Unfortunately I cannot live without that patch, i.e. can't try to reproduce with a pure vanilla vanilla kernel. 2) keepalived is 1.2.13 3) config uses "use_vmac" / "vmac_xmit_base", on multiple interfaces, i.e. MACVLAN interfaces on top of: 4) "normal" interfaces are both bridge-over-VLAN-over-LACP-bond-over-eth and ARP-bond-over-VLAN-over-eth 5) there is active conntracking including conntrackd (but excluding LVS state), LVS loadbalancing of some 15k pps, LVS sync, and heavy iptables use including ipset matching, going on. Just for completeness. Anybody got any idea what the root cause might be? best regards Patrick
