[ovs-dev] [PATCH 0/2] xlate: optimize dp flow action in case of error in multi-bridge setup

Zoltan Balogh zoltan.balogh at ericsson.com
Mon Oct 30 11:22:48 UTC 2017

    If several bridges are connected via patch ports or the packet is sent
    via tunnel and an error occurs during translation in a subsequent bridge,
    e.g. Ethernet header cannot be pushed to the packet due to lack of
    prerequisites in the second bridge and there were flow actions already
    translated in first bridge, it can happen that the installed datapath
    flow contains actions which are going to be performed for all received
    packets, however the packets are going to be dropped in the end. This
    is waste of processing resource.
    There are two points in the code where we can fix such datapath flows.
    1) During translation of OF actions on a bridge, we can store the last
    valid state of translated actions while iterating over the OF actions
    and revert to it in case of error. This can be performed in the
    do_xlate_actions() funtion.
    2) When all OF actions have been translated, there could be actions at
    the end of list of odp actions which are not needed to be executed.
    So, the list can be normalized at the end of xlate_actions().
    Let's have a config like this and inject L2 Ethernet packets into br0
    via LOCAL port:
          | LOCAL                  LOCAL                               LOCAL
        +-o-------+       +-------o-+       +---------+       +-------o-+
        |   br0   |       |   br1   |       |   br2   |       |   br3   |
        +-------o-+       +-o-----o-+       +-o-----o-+       +-o-------+
            p01 |       p10 |     | p12   p21 |     | p23   p32 |
                +-----------+     +-----------+     +-----------+
    netdev at ovs-netdev: hit:17754528 missed:57
                    br0 65534/1: (tap)
                    p01 10/none: (patch: peer=p10)
                    br1 65534/2: (tap)
                    p10 20/none: (patch: peer=p01)
                    p12 30/none: (patch: peer=p21)
                    br2 65534/3: (tap)
                    p21 40/none: (patch: peer=p12)
                    p23 50/none: (patch: peer=p32)
                    br3 65534/4: (tap)
                    p32 60/none: (patch: peer=p23)
    - OpenFlow rules:
    OFPST_FLOW reply (OF1.3) (xid=0x2):
    cookie=0x0, duration=15.601s, table=0, n_packets=20739, n_bytes=2032422,
    reset_counts in_port=LOCAL actions=dec_ttl,output:10
    OFPST_FLOW reply (OF1.3) (xid=0x2):
    cookie=0x0, duration=15.568s, table=0, n_packets=20739, n_bytes=2032422,
    reset_counts ip,in_port=20
    OFPST_FLOW reply (OF1.3) (xid=0x2):
    cookie=0x0, duration=15.536s, table=0, n_packets=20739, n_bytes=2032422,
    OFPST_FLOW reply (OF1.3) (xid=0x2):
    cookie=0x0, duration=15.502s, table=0, n_packets=0, n_bytes=0,
    reset_counts in_port=60 actions=LOCAL
    - Installed datapath flow:
    flow-dump from non-dpdk interfaces:
    ipv4(ttl=255,frag=no), packets:20738, bytes:2032324, used:5.176s,
    In this case, encap(ethernet) action cannot be performed, because of the
    packet is L2 Ethernet packet. By looking at the datapath flow, you can
    see, that setting MAC addresses and TTL at the end is actually not needed,
    since the packet is not transmitted anywhere after these actions.
    With this series, these actions are omitted. 

Zoltan Balogh (2):
  xlate: rollback to valid known state in case of ctx->error on
  xlate: normalize the actions after translation

 ofproto/ofproto-dpif-xlate.c | 178 +++++++++++++++++++++++++++++++++
 tests/ofproto-dpif.at        | 227 +++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 405 insertions(+)


More information about the dev mailing list