[ovs-dev] [PATCH 0/4] Adds capability to return packets back to the kernel

[Chris Luke]

Jesse, all,

This patchset provides a method to return packets to the kernel for
handling as though OVS never touched them.

To support this, we extend the datapath on kernels since 2.6.39 to be
able to pass packets back by providing a 'back_to_kernel' action which
causes the network input hook to return RX_HANDLER_PASS and userspace
packets to be re-submitted to the input queue.

The latter, the re-submission, can happen in one of two ways: Either
with a new function in the kernel that can deliver an skb to the lower
half of __netif_receive_skb_core() (see the patch at the HEAD of
https://github.com/chrisy/linux/tree/feature/netif-receive-finish )
or on other kernels without this function by submitting the packet to
the queue with netif_rx_ni().

The downside of using netif_rx_ni() is that the packet has already
traversed much of the input path and this risks duplication of action.
We mitigate this using a flag that causes some steps to be skipped, but
this is a hack (though the way we use this is cute in that the existing
implementation resets the flag for us so it seems there are no side
effects!)

Next, we add the 'back_to_kernel' action to OpenFlow which triggers a
datapath output action using the datapath 'back_to_kernel' action.

Suggestions for better names for 'back_to_kernel' welcome. :)

TODO:
- When using netif_rx_ni() to send packets back from userspace, we may
  be double-counting the first packet of a flow in the stats. This is
  since we have no way to avoid having the packet hit the rx_handler
  hook a second time and relying on it matching its flow.
- In the userspace datapath, look at whether AF_PACKET can reinsert
  packets at the right place in the input path, otherwise use a tap
  interface to do same.

Signed-off-by: Chris Luke <chris_luke at cable.comcast.com>
---
Revision 1:
- Alter kernel hook to be able to return packets to the kernel.
- Add tweak to vswitch to use this feature instead of a learning
   switch when it encounters OFPP_NORMAL.

Revision 2:  (supercedes patch 2606)
- Refactor the path taken through the datapath to remove one skb clone
   and lots of value returns.
- Add capability to insert a packet back to the kernel input queue.
- Add openflow action to send a packet back_to_kernel.
- Update tests/docs.

Revision 3:  (supercedes patch 2651)
- Break the patch into several;
  - Datapath back_to_kernel action.
  - OpenFlow action back_to_kernel. (requires datapath patch)
- Add tests for back_to_kernel method.
- Abandon the port_normal_mode and OFPP_NORMAL handling in vswitch patch.
- Alter datapath method from a special output port to a discrete action.
- skb header pointer "cleanup" has proven unnecessary.
- Alter how packets that went via userspace are reinput to the kernel
  (no more PACKET_LOOPBACK crud):
  - Hijack the tc_verd field, if we have it, in the skb to signal that
    the taps can be skipped. This works back to at least 2.6.36
    (which is when the current rx_handler logic was introduced.)
  - Proof of concept kernel patch to make this much more efficient.


Chris Luke (4):
  datapath: Action for returning packets to the kernel
  datapath: New kernel function to insert packets async
  lib,ofproto: Add 'back_to_kernel' OpenFlow action
  doc: Documentation updates for the back_to_kernel patches

 AUTHORS                       |    1 +
 FAQ                           |   87 ++++++++++++++++++++++++++++++-----
 NEWS                          |    2 +
 acinclude.m4                  |    3 ++
 datapath/actions.c            |   40 +++++++++++++++++
 datapath/datapath.c           |   23 ++++++++--
 datapath/datapath.h           |    9 +++-
 datapath/flow_netlink.c       |    6 ++-
 datapath/vport-netdev.c       |  100 ++++++++++++++++++++++++++++++++++++-----
 datapath/vport.c              |   26 ++++++++++-
 datapath/vport.h              |    4 +-
 include/linux/openvswitch.h   |    1 +
 include/openflow/nicira-ext.h |    1 +
 lib/dpif-netdev.c             |    1 +
 lib/dpif.c                    |    1 +
 lib/odp-execute.c             |    4 ++
 lib/odp-util.c                |   12 +++++
 lib/ofp-actions.c             |   30 +++++++++++++
 lib/ofp-actions.h             |    8 ++++
 lib/ofp-parse.c               |    4 ++
 lib/ofp-util.def              |    1 +
 ofproto/ofproto-dpif-xlate.c  |   27 +++++++++++
 tests/odp.at                  |    1 +
 tests/ofproto-dpif.at         |   15 +++++++
 tests/ovs-ofctl.at            |    2 +
 utilities/ovs-ofctl.8.in      |    7 +++
 26 files changed, 386 insertions(+), 30 deletions(-)

-- 
1.7.9.5