On Mon, Nov 16, 2020 at 02:28:44PM -0800, Jakub Kicinski wrote: > On Mon, 16 Nov 2020 23:18:15 +0100 Pablo Neira Ayuso wrote: > > On Sat, Nov 14, 2020 at 09:03:47AM -0800, Jakub Kicinski wrote: > > > On Sat, 14 Nov 2020 15:00:03 +0100 Tobias Waldekranz wrote: > > > > On Sat, Nov 14, 2020 at 12:59, Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx> wrote: > > > > > If any of the flowtable device goes down / removed, the entries are > > > > > removed from the flowtable. This means packets of existing flows are > > > > > pushed up back to classic bridge / forwarding path to re-evaluate the > > > > > fast path. > > > > > > > > > > For each new flow, the fast path that is selected freshly, so they use > > > > > the up-to-date FDB to select a new bridge port. > > > > > > > > > > Existing flows still follow the old path. The same happens with FIB > > > > > currently. > > > > > > > > > > It should be possible to explore purging entries in the flowtable that > > > > > are stale due to changes in the topology (either in FDB or FIB). > > > > > > > > > > What scenario do you have specifically in mind? Something like VM > > > > > migrates from one bridge port to another? > > > > > > Indeed, 2 VMs A and B, talking to each other, A is _outside_ the > > > system (reachable via eth0), B is inside (veth1). When A moves inside > > > and gets its veth. Neither B's veth1 not eth0 will change state, so > > > cache wouldn't get flushed, right? > > > > The flow tuple includes the input interface as part of the hash key, > > so packets will not match the existing entries in the flowtable after > > the topology update. > > To be clear - the input interface for B -> A traffic remains B. > So if B was talking to A before A moved it will keep hitting > the cached entry. Yes, Traffic for B -> A still hits the cached entry. > Are you saying A -> B traffic won't match so it will update the cache, > since conntrack flows are bi-directional? Yes, Traffic for A -> B won't match the flowtable entry, this will update the cache.
