On Mon, 30 Jan 2017 10:45:57 -0800 Roopa Prabhu <ro...@cumulusnetworks.com> wrote:
> On 1/30/17, 8:12 AM, David Ahern wrote: > > On 1/30/17 8:49 AM, Roopa Prabhu wrote: > >>> Single next hop delete will be around because IPv6 allows it -- and > >>> because IPv4 needs to support it. > >>> > >> understand single next hop delete for ipv6 will be around..and my only > >> point was to leave it around but not optimize for that case. > >> I don't think we should support single nexthop delete in ipv4 (I have not > >> seen a requirement for that)... ipv4 is good as it is right now. > >> the additional complexity is not needed. > >> > > IPv4 has a known bug -- delete a virtual interface in a multihop route and > > the entire route is deleted, including the nexthops for other devices. This > > does not happen for IPv6. > > > > Simple example of that bug: > > > > ip li add dummy1 type dummy > > ip li add dummy2 type dummy > > ip addr add dev dummy1 10.11.1.1/28 > > ip li set dummy1 up > > ip addr add dev dummy2 10.11.2.1/28 > > ip li set dummy2 up > > ip ro add 1.1.1.0/24 nexthop via 10.11.1.2 nexthop via 10.11.2.2 > > ip li del dummy2 > > > > --> the entire multipath route has been deleted. > > > > > > And, fixing this bug enables work to make IPv4 append to be sane -- > > appending a route should modify an existing route by adding the nexthop, > > not adding a new route that I believe can never actually be hit. > > > > Both cases mean modifying an IPv4 route -- adding or removing nexthops -- a > > capability that IPv6 allows so fixing this means closing another difference > > between the stacks. > > good point on the bug you point out. agree the bug needs to be fixed ...but > routing daemons react to this behavior pretty well...because they get a link > notification and a route notification. I was ok with fixing ipv6 to be > similar to ipv4...but I am not in favor of bringing in design choices that > ipv6 made into ipv4 :). > In all cases, in my experience with routes, the update of ecmp route as a > whole has always been ok (at-least not until now...maybe in the future > for new usecases) > > In the case of the bug you point out, can the user be notified of the > implicit update of the route in the kernel ...via replace flag or something ?. > regarding append..., ipv4 never really supported appending to an existing > route......even in the case of a non-ecmp routes. > append just dictates the order where the route is added IIRC (i maybe > mistaken here..its been long i tried it). My fear is that routing daemons already adapt to the funny semantics of multi-path routing in IPv4 vs IPv6 and therefore any change in semantics or flags risks breaking existing user space.
