[ovs-dev] [PATCH v1] netdev-linux: Remove device in netdev_shash when deleted

[Ben Pfaff]

On Wed, May 31, 2017 at 09:30:43AM +0800, fukaige wrote:
> Start a virtual machine with its backend tap device attached to a brought up linux bridge.
> If we delete the linux bridge when vm is still running, we'll get the following error when
> trying to create a ovs bridge with the same name.
> 
> The reason is that ovs-router subsystem add the linux bridge into netdev_shash, but does
> not remove it when the bridge is deleted in the situation. When the bridge is deleted, ovs
> will receive a RTM_DELLINK msg, take this chance to remove the bridge in netdev_shash.
> 
> ovs-vsctl: Error detected while setting up 'br-eth'.  See ovs-vswitchd log for details.
> 
> ovs-vswitchd log:
> 2017-05-11T03:45:25.293Z|00026|ofproto_dpif|INFO|system at ovs-system: Datapath supports recirculation
> 2017-05-11T03:45:25.293Z|00027|ofproto_dpif|INFO|system at ovs-system: MPLS label stack length probed as 1
> 2017-05-11T03:45:25.293Z|00028|ofproto_dpif|INFO|system at ovs-system: Datapath supports unique flow ids
> 2017-05-11T03:45:25.293Z|00029|ofproto_dpif|INFO|system at ovs-system: Datapath supports ct_state
> 2017-05-11T03:45:25.293Z|00030|ofproto_dpif|INFO|system at ovs-system: Datapath supports ct_zone
> 2017-05-11T03:45:25.293Z|00031|ofproto_dpif|INFO|system at ovs-system: Datapath supports ct_mark
> 2017-05-11T03:45:25.293Z|00032|ofproto_dpif|INFO|system at ovs-system: Datapath supports ct_label
> 2017-05-11T03:45:25.364Z|00001|ofproto_dpif_upcall(handler226)|INFO|received packet on unassociated datapath port 0
> 2017-05-11T03:45:25.368Z|00033|netdev_linux|WARN|ethtool command ETHTOOL_GFLAGS on network device br-eth failed: No such device
> 2017-05-11T03:45:25.368Z|00034|dpif|WARN|system at ovs-system: failed to add br-eth as port: No such device
> 2017-05-11T03:45:25.368Z|00035|bridge|INFO|bridge br-eth: using datapath ID 00002a51cf9f2841
> 2017-05-11T03:45:25.368Z|00036|connmgr|INFO|br-eth: added service controller "punix:/var/run/openvswitch/br-eth.mgmt"
> 
> Change-Id: Ib5ead59bc91453f83549da89937c0d3607e0385e
> Signed-off-by: fukaige <fukaige at huawei.com>

Thanks for identifying the problem and coming up with a fix.

The fix looks to me like it works at the wrong level of abstraction.
netdev-linux and tnl-port-map are not directly related, and I believe
that the same problem could arise with other kinds of netdevs and
tnl-port-map.  So, I think that it would be better if tnl-ports itself
registered a callback upon device destruction, for example via a call to
if_notifier_create() from tnl_port_map_init().

Thanks,

Ben.