On 08.11.2016 02:08, David Ahern wrote: > > Can anyone explain why host routes and anycast routes for IPv6 are added with > the device set to loopback versus the device with the address: > > local ::1 dev lo proto none metric 0 pref medium > local 2000:1:: dev lo proto none metric 0 pref medium > local 2000:1::3 dev lo proto none metric 0 pref medium > local 2100:2:: dev lo proto none metric 0 pref medium > local 2100:2::3 dev lo proto none metric 0 pref medium
Does it really matter? For global valid unicast addresses we still implement the weak model. Thus the interface does not matter at all. > This behavior differs from IPv4 where host routes use the device with the > address: > > broadcast 10.1.1.0 dev eth0 proto kernel scope link src 10.1.1.3 > local 10.1.1.3 dev eth0 proto kernel scope host src 10.1.1.3 > broadcast 10.1.1.255 dev eth0 proto kernel scope link src 10.1.1.3 > broadcast 10.100.2.0 dev eth2 proto kernel scope link src 10.100.2.3 > local 10.100.2.3 dev eth2 proto kernel scope host src 10.100.2.3 > broadcast 10.100.2.255 dev eth2 proto kernel scope link src 10.100.2.3 > > The use of loopback pre-dates the git history, so wondering if someone > recalls the reason why. We would like to change that to make it consistent > with IPv4 - with a sysctl to maintain backwards compatibility. A sysctl for that sounds like a really bad idea. Internally the sysctl will change the reference counting of interfaces and routes towards each other, have small but difficult to find semantically changes inside the kernel, just for switchting the interface in iproute/netlink dumps? If there a good reasons (which can very well be) to switch to have the interface with the address in the routes, we should switch without providing the backwards compatibility sysctl. Bye, Hannes
