netdev-dpdk: Add support for vHost dequeue zero copy (experimental)

Zero copy is disabled by default. To enable it, set the 'dq-zero-copy'
option to 'true' when configuring the Interface:

ovs-vsctl set Interface dpdkvhostuserclient0
options:vhost-server-path=/tmp/dpdkvhostuserclient0
options:dq-zero-copy=true

When packets from a vHost device with zero copy enabled are destined for
a single 'dpdk' port, the number of tx descriptors on that 'dpdk' port
must be set to a smaller value. 128 is recommended. This can be achieved
like so:

ovs-vsctl set Interface dpdkport options:n_txq_desc=128

Note: The sum of the tx descriptors of all 'dpdk' ports the VM will send
to should not exceed 128. Due to this requirement, the feature is
considered 'experimental'.

Testing of the patch showed a ~8% improvement when switching 512B
packets between vHost devices on different VMs on the same host when
zero copy was enabled on the transmitting device.

Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Acked-by: Ilya Maximets <i.maximets@samsung.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>

2 files changed, 75 insertions, 0 deletions

diff --git a/Documentation/intro/install/dpdk.rst b/Documentation/intro/install/dpdk.rst
index 8d5ee562d..ed358d577 100644
--- a/Documentation/intro/install/dpdk.rst
+++ b/Documentation/intro/install/dpdk.rst
@@ -518,6 +518,8 @@ The above command sets the number of rx queues for DPDK physical interface.
 The rx queues are assigned to pmd threads on the same NUMA node in a
 round-robin fashion.
 
+.. _dpdk-queues-sizes:
+
 DPDK Physical Port Queue Sizes
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
diff --git a/Documentation/topics/dpdk/vhost-user.rst b/Documentation/topics/dpdk/vhost-user.rst
index 8447e2dee..95517a676 100644
--- a/Documentation/topics/dpdk/vhost-user.rst
+++ b/Documentation/topics/dpdk/vhost-user.rst
@@ -458,3 +458,76 @@ Sample XML
     </domain>
 
 .. _QEMU documentation: http://git.qemu-project.org/?p=qemu.git;a=blob;f=docs/specs/vhost-user.txt;h=7890d7169;hb=HEAD
+
+vhost-user Dequeue Zero Copy (experimental)
+-------------------------------------------
+
+Normally when dequeuing a packet from a vHost User device, a memcpy operation
+must be used to copy that packet from guest address space to host address
+space. This memcpy can be removed by enabling dequeue zero-copy like so::
+
+    $ ovs-vsctl add-port br0 dpdkvhostuserclient0 -- set Interface \
+        dpdkvhostuserclient0 type=dpdkvhostuserclient \
+        options:vhost-server-path=/tmp/dpdkvhostclient0 \
+        options:dq-zero-copy=true
+
+With this feature enabled, a reference (pointer) to the packet is passed to
+the host, instead of a copy of the packet. Removing this memcpy can give a
+performance improvement for some use cases, for example switching large packets
+between different VMs. However additional packet loss may be observed.
+
+Note that the feature is disabled by default and must be explicitly enabled
+by setting the ``dq-zero-copy`` option to ``true`` while specifying the
+``vhost-server-path`` option as above. If you wish to split out the command
+into multiple commands as below, ensure ``dq-zero-copy`` is set before
+``vhost-server-path``::
+
+    $ ovs-vsctl set Interface dpdkvhostuserclient0 options:dq-zero-copy=true
+    $ ovs-vsctl set Interface dpdkvhostuserclient0 \
+        options:vhost-server-path=/tmp/dpdkvhostclient0
+
+The feature is only available to ``dpdkvhostuserclient`` port types.
+
+A limitation exists whereby if packets from a vHost port with
+``dq-zero-copy=true`` are destined for a ``dpdk`` type port, the number of tx
+descriptors (``n_txq_desc``) for that port must be reduced to a smaller number,
+128 being the recommended value. This can be achieved by issuing the following
+command::
+
+    $ ovs-vsctl set Interface dpdkport options:n_txq_desc=128
+
+Note: The sum of the tx descriptors of all ``dpdk`` ports the VM will send to
+should not exceed 128. For example, in case of a bond over two physical ports
+in balance-tcp mode, one must divide 128 by the number of links in the bond.
+
+Refer to :ref:`dpdk-queues-sizes` for more information.
+
+The reason for this limitation is due to how the zero copy functionality is
+implemented. The vHost device's 'tx used vring', a virtio structure used for
+tracking used ie. sent descriptors, will only be updated when the NIC frees
+the corresponding mbuf. If we don't free the mbufs frequently enough, that
+vring will be starved and packets will no longer be processed. One way to
+ensure we don't encounter this scenario, is to configure ``n_txq_desc`` to a
+small enough number such that the 'mbuf free threshold' for the NIC will be hit
+more often and thus free mbufs more frequently. The value of 128 is suggested,
+but values of 64 and 256 have been tested and verified to work too, with
+differing performance characteristics. A value of 512 can be used too, if the
+virtio queue size in the guest is increased to 1024 (available to configure in
+QEMU versions v2.10 and greater). This value can be set like so::
+
+    $ qemu-system-x86_64 ... -chardev socket,id=char1,path=<sockpath>,server
+      -netdev type=vhost-user,id=mynet1,chardev=char1,vhostforce
+      -device virtio-net-pci,mac=00:00:00:00:00:01,netdev=mynet1,
+      tx_queue_size=1024
+
+Because of this limitation, this feature is considered 'experimental'.
+
+The feature currently does not fully work with QEMU >= v2.7 due to a bug in
+DPDK which will be addressed in an upcoming release. The patch to fix this
+issue can be found on
+`Patchwork
+<http://dpdk.org/dev/patchwork/patch/32198/>`__
+
+Further information can be found in the
+`DPDK documentation
+<http://dpdk.readthedocs.io/en/v17.05/prog_guide/vhost_lib.html>`__