Eyal Birger <eyal.bir...@gmail.com> wrote: Sorry for taking so long to respond.
> On Tue, 6 Feb 2018 14:15:09 +0100 > Florian Westphal <f...@strlen.de> wrote: > > > Steffen Klassert <steffen.klass...@secunet.com> wrote: > > > I gave the patch a quick try, but still I get this: > > > > > > unregister_netdevice: waiting for dummy1 to become free. Usage > > > count = 2 > > > > Was that with Eyals setup or the bridge one I posted? > > > > If it was Eyals setup, its possible the patch missed hookup > > to whatever tunnel infra is used (the setup I used has ipip tunnel, > > everything is ipv4). > > > > Thanks! > > Indeed the setup I'm testing uses ip6_tunnel. > I have tested a fix in the spirit of the patch and it looks valid > for ip6_tunnel as well. > > It looks though that this change would need to be added to any tunnel > device using dst_cache (vxlan, geneve, gre, ...). Yes. Meanwhile I tested your patch and it works for me too. As your patch is shorter and ipv4/ipv6 seem to take care of refcount put just fine I think your patch is the right way to go. The xfrm_dst size incrase isn't much of a big deal, there is ample of padding at the end so it will still be allocated from same slab. We could reduce num_pols and num_xfrms to u8, which creates a 16 bit hole, then store the cpu number instead of a list pointer. This would limit growth to 16 instead of 24. But, as I said, i do not think its a big deal. > I'm wondering - non-xfrm dsts are already correctly invalidated, > so do you think it makes sense to invalidate caches for devices that > have no xfrm dsts? or maybe I didn't understand the suggestion? See above, I think your patch is the way to go.
