numam-dpdk/doc/guides/rawdevs/ntb.rst

..  SPDX-License-Identifier: BSD-3-Clause
    Copyright(c) 2018 Intel Corporation.

NTB Rawdev Driver
=================

The ``ntb`` rawdev driver provides a non-transparent bridge between two
separate hosts so that they can communicate with each other. Thus, many
user cases can benefit from this, such as fault tolerance and visual
acceleration.

This PMD allows two hosts to handshake for device start and stop, memory
allocation for the peer to access and read/write allocated memory from peer.
Also, the PMD allows to use doorbell registers to notify the peer and share
some information by using scratchpad registers.

BIOS setting on Intel Skylake
-----------------------------

Intel Non-transparent Bridge needs special BIOS setting. Since the PMD only
supports Intel Skylake platform, introduce BIOS setting here. The referencce
is https://www.intel.com/content/dam/support/us/en/documents/server-products/Intel_Xeon_Processor_Scalable_Family_BIOS_User_Guide.pdf

- Set the needed PCIe port as NTB to NTB mode on both hosts.
- Enable NTB bars and set bar size of bar 23 and bar 45 as 12-29 (2K-512M)
  on both hosts. Note that bar size on both hosts should be the same.
- Disable split bars for both hosts.
- Set crosslink control override as DSD/USP on one host, USD/DSP on
  another host.
- Disable PCIe PII SSC (Spread Spectrum Clocking) for both hosts. This
  is a hardware requirement.

Build Options
-------------

- ``CONFIG_RTE_LIBRTE_PMD_NTB_RAWDEV`` (default ``y``)

   Toggle compilation of the ``ntb`` driver.

Device Setup
------------

The Intel NTB devices need to be bound to a DPDK-supported kernel driver
to use, i.e. igb_uio, vfio. The ``dpdk-devbind.py`` script can be used to
show devices status and to bind them to a suitable kernel driver. They will
appear under the category of "Misc (rawdev) devices".

Prerequisites
-------------

NTB PMD needs kernel PCI driver to support write combining (WC) to get
better performance. The difference will be more than 10 times.
To enable WC, there are 2 ways.

- Insert igb_uio with ``wc_activate=1`` flag if use igb_uio driver.

.. code-block:: console

  insmod igb_uio.ko wc_activate=1

- Enable WC for NTB device's Bar 2 and Bar 4 (Mapped memory) manually.
  The reference is https://www.kernel.org/doc/html/latest/x86/mtrr.html
  Get bar base address using ``lspci -vvv -s ae:00.0 | grep Region``.

.. code-block:: console

  # lspci -vvv -s ae:00.0 | grep Region
  Region 0: Memory at 39bfe0000000 (64-bit, prefetchable) [size=64K]
  Region 2: Memory at 39bfa0000000 (64-bit, prefetchable) [size=512M]
  Region 4: Memory at 39bfc0000000 (64-bit, prefetchable) [size=512M]

Using the following command to enable WC.

.. code-block:: console

  echo "base=0x39bfa0000000 size=0x20000000 type=write-combining" >> /proc/mtrr
  echo "base=0x39bfc0000000 size=0x20000000 type=write-combining" >> /proc/mtrr

And the results:

.. code-block:: console

  # cat /proc/mtrr
  reg00: base=0x000000000 (    0MB), size= 2048MB, count=1: write-back
  reg01: base=0x07f000000 ( 2032MB), size=   16MB, count=1: uncachable
  reg02: base=0x39bfa0000000 (60553728MB), size=  512MB, count=1: write-combining
  reg03: base=0x39bfc0000000 (60554240MB), size=  512MB, count=1: write-combining

To disable WC for these regions, using the following.

.. code-block:: console

     echo "disable=2" >> /proc/mtrr
     echo "disable=3" >> /proc/mtrr

Ring Layout
-----------

Since read/write remote system's memory are through PCI bus, remote read
is much more expensive than remote write. Thus, the enqueue and dequeue
based on ntb ring should avoid remote read. The ring layout for ntb is
like the following:

- Ring Format::

   desc_ring:

      0               16                                              64
      +---------------------------------------------------------------+
      |                        buffer address                         |
      +---------------+-----------------------------------------------+
      | buffer length |                      resv                     |
      +---------------+-----------------------------------------------+

   used_ring:

      0               16              32
      +---------------+---------------+
      | packet length |     flags     |
      +---------------+---------------+

- Ring Layout::

      +------------------------+   +------------------------+
      | used_ring              |   | desc_ring              |
      | +---+                  |   | +---+                  |
      | |   |                  |   | |   |                  |
      | +---+      +--------+  |   | +---+                  |
      | |   | ---> | buffer | <+---+-|   |                  |
      | +---+      +--------+  |   | +---+                  |
      | |   |                  |   | |   |                  |
      | +---+                  |   | +---+                  |
      |  ...                   |   |  ...                   |
      |                        |   |                        |
      |            +---------+ |   |            +---------+ |
      |            | tx_tail | |   |            | rx_tail | |
      | System A   +---------+ |   | System B   +---------+ |
      +------------------------+   +------------------------+
                    <---------traffic---------

- Enqueue and Dequeue
  Based on this ring layout, enqueue reads rx_tail to get how many free
  buffers and writes used_ring and tx_tail to tell the peer which buffers
  are filled with data.
  And dequeue reads tx_tail to get how many packets are arrived, and
  writes desc_ring and rx_tail to tell the peer about the new allocated
  buffers.
  So in this way, only remote write happens and remote read can be avoid
  to get better performance.

Limitation
----------

- This PMD only supports Intel Skylake platform.