On Thu, Oct 05, 2000 at 10:08:18AM -0400, Arni Raghu wrote: > What are the timeouts like...? > > Like if the linux router decides to route packets to address1 over interface > eth1 at time t1, is there any kind of time t2 later when the router will > start sending the packets over the other interface eth2 for the same > address1.. i.e how are sessions defined here..?? > > I know I should be looking at the source..;) The only source of failover is an explicit ifconfig down currently. You could e.g. use a OSPF daemon to let it do the failover, although I unfortunately don't know an OSPF daemon that would support linux multipath routing. You could always write your own daemon for it. There is also another path. Linux already maintains a timer that could tell it easily when a link makes no progress anymore: the ARP neighbour state machine which does regular unicast ARP probing and maintains states. It is not used for multipath routing though, because there would be no way to recover. The neighbour state machine only is active when it has packets in its queue, when multipath routing would stop sending packets over an device after a timeout it would never recover. There is a ugly trick to work around it. Patch the multipoint routing selection mechanism to take neighbour states into account (that's very easy) Use a user space daemon that regularly uses a different (non multiple nexthop) route with the same IP (using policy routing) to ping the nexthops of your multipoint routes. This way the neighbour state machine would always maintain neighbour states and the nexthop routing would recover links after failures. Using a OSPF daemon is probably easier though than abusing ARP like this, because it was designed for these things. Just someone has to do the work of teaching one of the Linux OSPF implementations (e.g. bird or zebra) about multiple nexthop routes. -Andi - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org
