summaryrefslogtreecommitdiffstats
path: root/src/vnet/ip/punt.rst
diff options
context:
space:
mode:
authorNathan Skrzypczak <nathan.skrzypczak@gmail.com>2019-07-29 11:18:05 +0200
committerDave Barach <openvpp@barachs.net>2019-08-01 18:01:57 +0000
commit55c68c9521d98005ce850ee54a40c7579d88928b (patch)
treeb5513b058e87e9914c13e5f08df5787f2b595fa8 /src/vnet/ip/punt.rst
parent30aaf97a90230d68c0f2736b0a026e07b06e7e32 (diff)
docs: Small fixes and ordering
* deactivtates the TODOs on doxygen (were empty) * This move punt.md to readthedocs (should be the new place for dev doc ?) * Makes Handoff queue demo plugin a child of dev doc in doxygen Type: fix Change-Id: I1f0476a911b35208212af8dd608bc76160efd22a Signed-off-by: Nathan Skrzypczak <nathan.skrzypczak@gmail.com>
Diffstat (limited to 'src/vnet/ip/punt.rst')
-rw-r--r--src/vnet/ip/punt.rst81
1 files changed, 81 insertions, 0 deletions
diff --git a/src/vnet/ip/punt.rst b/src/vnet/ip/punt.rst
new file mode 100644
index 00000000000..3f1a9d4a844
--- /dev/null
+++ b/src/vnet/ip/punt.rst
@@ -0,0 +1,81 @@
+.. _dev_punt:
+
+.. toctree::
+
+Punting Packets
+===============
+
+Overview
+________
+
+To 'punt' can mean different things to different people. In VPP the
+data-plane punts when a packet cannot be handled by any further
+nodes. Punt differs from drop, in that VPP is giving other elements of
+the system the opportunity to handle this packet.
+
+A popular meaning of punt is to send packets to the user/control-plane.
+This is specific option of the more general case above, where VPP is
+handing the packet to the control-plane for further prosessing.
+
+The Punt Infrastructure
+_______________________
+
+Exception packets are those that a given node cannot handle via normal
+mecahnisms.
+Punting of exception packets is handled via the VLIB 'punt
+infra'. There are two types of nodes; sources and sinks. Sources
+allocate a punt 'reason' from the infrastructre and load time. When
+they encouter an exception during switch time it will tag the packet
+with the reason and ship the packet of the the punt-dispatch node. A
+sink will register with the punt infra at load time so it can receive
+packets that were punted for that reason. If no sinks are registerd
+for a given reason the packet is dropped, if multiple sinks register
+the packets are replicated.
+
+This mechanism allows us to extend the system to deal with packets
+that the source node would otherise drop.
+
+
+Punting to the Control Plane
+____________________________
+
+Active Punt
+-----------
+
+The user/control-plane specifies that this is the type of packet I
+want to receive and this is where I want it sent.
+
+Currently there exists 3 ways to describe how to match/classify the
+packets to be punted:
+
+1) a matching UDP port
+2) a matching IP protocol (i.e. OSPF)
+3) a matching punt excpetion reason (see above)
+
+Depending on the type/classification of the packet to be punted, that
+active punt will register itself into the VLIB graph to receive those
+packets. For example, if it's a packet matching a UDP port then it
+will hook into the UDP port dispatch functions; udp_register_port().
+
+There exists only one sink for passive punt, a unix domain socket. But
+more work is underway in this area.
+
+see the API in: vnet/ip/punt.api
+
+
+
+Passive Punt
+------------
+
+VPP input packet processing can be described as a series of
+classifiers. For example, a sequence of input classifications could
+be, is it IP? is it for-us? is it UDP? is it a known UDP-port? If at
+some point in this pipline VPP has no further classifications to make,
+then the packet can be punted, which means sent to ipX-punt node. This
+is described as passive since the control-plane is thus receiving
+every packet that VPP does not itself handle.
+For passive punt the user can specify where the packets should be
+sent and whether/how they should be policed/rate-limited.
+
+see the API in: vnet/ip/ip.api
+