On Mon, Oct 1, 2012 at 8:52 AM, Kyle Mestery <kmest...@cisco.com> wrote: > Move struct ovs_key_ipv4_tunnel out of struct > sw_flow_key and into struct sw_flow. This allows it to > "float" and be used for matching only when needed. > Modify the matching code in ovs_flow_tbl_lookup() to > match on the tunnel header if it's set.
This is a little bit different from what I was originally envisioning but I think we can combine them to get the best of both worlds. As it currently exists, there are a couple of things that worry me a little bit: * While the tunnel information is included in the lookup comparison, it's not in the hash. While it's probably not a likely situation that many different tunnel endpoints are sending the same flow, it is possible. It's also possible that a malicious user could use it to unbalance the hash table. * Ideally, I'd like to keep optimizations like this internal to flow.c. There's a little more code change in datapath.c than I'd like to see. Here's my idea: * Keep tun_key at the very beginning of the struct sw_flow_key and populate it like usual. * Create a new struct containing hashing/comparison parameters. This is basically just data passed between flow_extract()/ovs_flow_from_nlattrs() and ovs_flow_hash()/ovs_flow_tbl_lookup(). Initially, it would just contain key_len. * Add a key_start offset to the new comparison parameters struct, which would point to either phy if there is no tunnel or otherwise tun_key. So essentially we make the flow key variable length in both directions. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
