Re: [RFC net-next 0/2] net/sched: cls_flower, act_mirred: VXLAN redirect using TC

On Mon, Aug 15, 2016 at 06:41:14AM -0400, Jamal Hadi Salim wrote:

On 16-08-15 06:24 AM, Shmulik Ladkani wrote:

quoted

On Mon, 15 Aug 2016 06:08:10 -0400, jhs@mojatatu.com wrote:

quoted

quoted

Assuming $VXLAN is actually not a linux netdev of type vxlan?
then the action does vxlan encap redirect sends it to the $VXLAN
dev with encapsulation in place.

I assume Amir refers to vxlan netdev in VXLAN_F_COLLECT_METADATA mode,
using the tun_info metadata found in skb_metadata_dst.
The action is supposed to assign the tun metadata.

I see - so you let the vxlan netdev do the encap?
Would it still scale to a _very large_ number of tunnels?
How many netdevs are you going to use? I am assuming you will hit
a nasty lock somewhere(qdisc?) if you use only one.

Having a netdev per tunnel is problematic in its memory use [1].
User can take each of the approaches. Can have a shared netdev, but will
have some contention on the qdisc lock, or create a vxlan dev per VNI
and increase memory use.
When offloading will be added, shared netdev will enjoy all worlds - low
memory use and no lock contention.


[1] - http://www.netdevconf.org/1.1/proceedings/slides/ahern-aleksandrov-prabhu-scaling-network-cumulus.pdf

cheers,
jamal