clang: Add annotations for thread safety check.

[sliver-openvswitch.git] / INSTALL

      How to Install Open vSwitch on Linux, FreeBSD and NetBSD
      ========================================================

This document describes how to build and install Open vSwitch on a
generic Linux, FreeBSD, or NetBSD host. For specifics around installation
on a specific platform, please see one of these files:

    - INSTALL.Debian
    - INSTALL.Fedora
    - INSTALL.RHEL
    - INSTALL.XenServer

Build Requirements
------------------

To compile the userspace programs in the Open vSwitch distribution,
you will need the following software:

    - GNU make.

    - The GNU C compiler.  We generally test with version 4.1, 4.2, or
      4.3.

    - libssl, from OpenSSL, is optional but recommended if you plan to
      connect the Open vSwitch to an OpenFlow controller.  libssl is
      required to establish confidentiality and authenticity in the
      connections from an Open vSwitch to an OpenFlow controller.  If
      libssl is installed, then Open vSwitch will automatically build
      with support for it.

    - clang, from LLVM, is optional. It provides useful static semantic
      analyzer and thread-safety check. clang version must be 3.4 or
      later. For Ubuntu, there are nightly built packages available on
      clang's website.

To compile the kernel module on Linux, you must also install the
following.  If you cannot build or install the kernel module, you may
use the userspace-only implementation, at a cost in performance.  The
userspace implementation may also lack some features.  Refer to
INSTALL.userspace for more information.

    - A supported Linux kernel version.  Please refer to README for a
      list of supported versions.

      The Open vSwitch datapath requires bridging support
      (CONFIG_BRIDGE) to be built as a kernel module.  (This is common
      in kernels provided by Linux distributions.)  The bridge module
      must not be loaded or in use.  If the bridge module is running
      (check with "lsmod | grep bridge"), you must remove it ("rmmod
      bridge") before starting the datapath.

      For optional support of ingress policing, you must enable kernel
      configuration options NET_CLS_BASIC, NET_SCH_INGRESS, and
      NET_ACT_POLICE, either built-in or as modules.  (NET_CLS_POLICE is
      obsolete and not needed.)

      To use GRE tunneling on Linux 2.6.37 or newer, kernel support
      for GRE must be compiled in or available as a module
      (CONFIG_NET_IPGRE_DEMUX).

      To configure HTB or HFSC quality of service with Open vSwitch,
      you must enable the respective configuration options.

      To use Open vSwitch support for TAP devices, you must enable
      CONFIG_TUN.

    - To build a kernel module, you need the same version of GCC that
      was used to build that kernel.

    - A kernel build directory corresponding to the Linux kernel image
      the module is to run on.  Under Debian and Ubuntu, for example,
      each linux-image package containing a kernel binary has a
      corresponding linux-headers package with the required build
      infrastructure.

If you are working from a Git tree or snapshot (instead of from a
distribution tarball), or if you modify the Open vSwitch build system
or the database schema, you will also need the following software:

    - Autoconf version 2.64 or later.

    - Automake version 1.10 or later.

    - Python 2.x, for x >= 4.

If you modify the ovsdbmonitor tool, then you will also need the
following:

    - pyuic4 from PyQt4 (http://www.riverbankcomputing.co.uk).

To run the unit tests, you also need:

    - Perl.  Version 5.10.1 is known to work.  Earlier versions should
      also work.

If you modify the vswitchd database schema, then the E-R diagram in
the ovs-vswitchd.conf.db(5) manpage will be updated properly only if
you have the following:

    - "dot" from graphviz (http://www.graphviz.org/).

    - Perl.  Version 5.10.1 is known to work.  Earlier versions should
      also work.

    - Python 2.x, for x >= 4.

If you are going to extensively modify Open vSwitch, please consider
installing the following to obtain better warnings:

    - "sparse" version 0.4.4 or later
      (http://www.kernel.org/pub/software/devel/sparse/dist/).

    - GNU make.

    - clang, version 3.4 or later

Also, you may find the ovs-dev script found in utilities/ovs-dev.py useful.

Installation Requirements
-------------------------

The machine on which Open vSwitch is to be installed must have the
following software:

    - libc compatible with the libc used for build.

    - libssl compatible with the libssl used for build, if OpenSSL was
      used for the build.

    - On Linux, the same kernel version configured as part of the build.

    - For optional support of ingress policing on Linux, the "tc" program
      from iproute2 (part of all major distributions and available at
      http://www.linux-foundation.org/en/Net:Iproute2).

On Linux you should ensure that /dev/urandom exists.  To support TAP
devices, you must also ensure that /dev/net/tun exists.

To run the ovsdbmonitor tool, the machine must also have the following
software:

    - Python 2.x, for x >= 4.

    - Python Twisted Conch.

    - Python JSON.

    - PySide or PyQt4.

    - Python Zope interface module.

(On Debian "lenny" the above can be installed with "apt-get install
python-json python-qt4 python-zopeinterface python-twisted-conch".)

Building and Installing Open vSwitch for Linux, FreeBSD or NetBSD
=================================================================

Once you have installed all the prerequisites listed above in the Base
Prerequisites section, follow the procedure below to build.

1. If you pulled the sources directly from an Open vSwitch Git tree,
   run boot.sh in the top source directory:

      % ./boot.sh

2. In the top source directory, configure the package by running the
   configure script.  You can usually invoke configure without any
   arguments:

      % ./configure

   By default all files are installed under /usr/local.  If you want
   to install into, e.g., /usr and /var instead of /usr/local and
   /usr/local/var, add options as shown here:

      % ./configure --prefix=/usr --localstatedir=/var

   To use a specific C compiler for compiling Open vSwitch user
   programs, also specify it on the configure command line, like so:

      % ./configure CC=gcc-4.2

   To use 'clang' compiler:

      % ./configure CC=clang

   To build the Linux kernel module, so that you can run the
   kernel-based switch, pass the location of the kernel build
   directory on --with-linux.  For example, to build for a running
   instance of Linux:

      % ./configure --with-linux=/lib/modules/`uname -r`/build

   If you wish to build the kernel module for an architecture other
   than the architecture of the machine used for the build, you may
   specify the kernel architecture string using the KARCH variable
   when invoking the configure script.  For example, to build for MIPS
   with Linux:

      % ./configure --with-linux=/path/to/linux KARCH=mips

   The configure script accepts a number of other options and honors
   additional environment variables.  For a full list, invoke
   configure with the --help option.

3. Run GNU make in the top source directory, e.g.:

      % make

   or if GNU make is installed as "gmake":

      % gmake

   For improved warnings if you installed "sparse" (see
   "Prerequisites"), add C=1 to the command line.

4. Consider running the testsuite.  Refer to "Running the Testsuite"
   below, for instructions.

5. Become root by running "su" or another program.

6. Run "make install" to install the executables and manpages into the
   running system, by default under /usr/local.

7. If you built kernel modules, you may install and load them, e.g.:

      % make modules_install
      % /sbin/modprobe openvswitch 

   To verify that the modules have been loaded, run "/sbin/lsmod" and
   check that openvswitch is listed.

   If the "modprobe" operation fails, look at the last few kernel log
   messages (e.g. with "dmesg | tail"):

      - The message "openvswitch: exports duplicate symbol
        br_should_route_hook (owned by bridge)" means that the bridge
        module is loaded.  Run "/sbin/rmmod bridge" to remove it.

        If "/sbin/rmmod bridge" fails with "ERROR: Module bridge does
        not exist in /proc/modules", then the bridge is compiled into
        the kernel, rather than as a module.  Open vSwitch does not
        support this configuration (see "Build Requirements", above).

      - The message "openvswitch: exports duplicate symbol
        dp_ioctl_hook (owned by ofdatapath)" means that the ofdatapath
        module from the OpenFlow reference implementation is loaded.
        Run "/sbin/rmmod ofdatapath" to remove it.  (You might have to
        delete any existing datapaths beforehand, using the "dpctl"
        program included with the OpenFlow reference implementation.
        "ovs-dpctl" will not work.)

      - Otherwise, the most likely problem is that Open vSwitch was
        built for a kernel different from the one into which you are
        trying to load it.  Run "modinfo" on openvswitch.ko and on
        a module built for the running kernel, e.g.:

           % /sbin/modinfo openvswitch.ko
           % /sbin/modinfo /lib/modules/`uname -r`/kernel/net/bridge/bridge.ko

        Compare the "vermagic" lines output by the two commands.  If
        they differ, then Open vSwitch was built for the wrong kernel.

      - If you decide to report a bug or ask a question related to
        module loading, please include the output from the "dmesg" and
        "modinfo" commands mentioned above.

   There is an optional module parameter to openvswitch.ko called
   vlan_tso that enables TCP segmentation offload over VLANs on NICs
   that support it. Many drivers do not expose support for TSO on VLANs
   in a way that Open vSwitch can use but there is no way to detect
   whether this is the case. If you know that your particular driver can
   handle it (for example by testing sending large TCP packets over VLANs)
   then passing in a value of 1 may improve performance. Modules built for
   Linux kernels 2.6.37 and later, as well as specially patched versions
   of earlier kernels, do not need this and do not have this parameter. If
   you do not understand what this means or do not know if your driver
   will work, do not set this.

8. Initialize the configuration database using ovsdb-tool, e.g.:

      % mkdir -p /usr/local/etc/openvswitch
      % ovsdb-tool create /usr/local/etc/openvswitch/conf.db vswitchd/vswitch.ovsschema

Startup
=======

Before starting ovs-vswitchd itself, you need to start its
configuration database, ovsdb-server.  Each machine on which Open
vSwitch is installed should run its own copy of ovsdb-server.
Configure it to use the database you created during step 7 of
installation, above, to listen on a Unix domain socket, to connect to
any managers specified in the database itself, and to use the SSL
configuration in the database:

      % ovsdb-server --remote=punix:/usr/local/var/run/openvswitch/db.sock \
                     --remote=db:Open_vSwitch,Open_vSwitch,manager_options \
                     --private-key=db:Open_vSwitch,SSL,private_key \
                     --certificate=db:Open_vSwitch,SSL,certificate \
                     --bootstrap-ca-cert=db:Open_vSwitch,SSL,ca_cert \
                     --pidfile --detach

(If you built Open vSwitch without SSL support, then omit
--private-key, --certificate, and --bootstrap-ca-cert.)

Then initialize the database using ovs-vsctl.  This is only
necessary the first time after you create the database with
ovsdb-tool (but running it at any time is harmless):

      % ovs-vsctl --no-wait init

Then start the main Open vSwitch daemon, telling it to connect to the
same Unix domain socket:

      % ovs-vswitchd --pidfile --detach

Now you may use ovs-vsctl to set up bridges and other Open vSwitch
features.  For example, to create a bridge named br0 and add ports
eth0 and vif1.0 to it:

      % ovs-vsctl add-br br0
      % ovs-vsctl add-port br0 eth0
      % ovs-vsctl add-port br0 vif1.0

Please refer to ovs-vsctl(8) for more details.

Upgrading
=========

When you upgrade Open vSwitch from one version to another, you should
also upgrade the database schema:

1. Stop the Open vSwitch daemons, e.g.:

      % kill `cd /usr/local/var/run/openvswitch && cat ovsdb-server.pid ovs-vswitchd.pid`

2. Install the new Open vSwitch release.

3. Upgrade the database, in one of the following two ways:

      - If there is no important data in your database, then you may
        delete the database file and recreate it with ovsdb-tool,
        following the instructions under "Building and Installing Open
        vSwitch for Linux, FreeBSD or NetBSD".

      - If you want to preserve the contents of your database, back it
        up first, then use "ovsdb-tool convert" to upgrade it, e.g.:

        % ovsdb-tool convert /usr/local/etc/openvswitch/conf.db vswitchd/vswitch.ovsschema

4. Start the Open vSwitch daemons as described under "Building and
   Installing Open vSwitch for Linux, FreeBSD or NetBSD" above.

Hot Upgrading
=============
Upgrading Open vSwitch from one version to the next version with minimum
disruption of traffic going through the system that is using that Open vSwitch
needs some considerations:

1. If the upgrade only involves upgrading the userspace utilities and daemons
of Open vSwitch, make sure that the new userspace version is compatible with
the previously loaded kernel module.

2. An upgrade of userspace daemons means that they have to be restarted.
Restarting the daemons means that the Openflow flows in the ovs-vswitchd daemon
will be lost.  One way to restore the flows is to let the controller
re-populate it.  Another way is to save the previous flows using a utility
like ovs-ofctl and then re-add them after the restart. Restoring the old flows
is accurate only if the new Open vSwitch interfaces retain the old 'ofport'
values.

3. When the new userspace daemons get restarted, they automatically flush
the old flows setup in the kernel. This can be expensive if there are hundreds
of new flows that are entering the kernel but userspace daemons are busy
setting up new userspace flows from either the controller or an utility like
ovs-ofctl. Open vSwitch database provides an option to solve this problem
through the other_config:flow-restore-wait column of the Open_vSwitch table.
Refer to the ovs-vswitchd.conf.db(5) manpage for details.

4. If the upgrade also involves upgrading the kernel module, the old kernel
module needs to be unloaded and the new kernel module should be loaded. This
means that the kernel network devices belonging to Open vSwitch is recreated
and the kernel flows are lost. The downtime of the traffic can be reduced
if the userspace daemons are restarted immediately and the userspace flows
are restored as soon as possible.

The ovs-ctl utility's "restart" function only restarts the userspace daemons,
makes sure that the 'ofport' values remain consistent across restarts, restores
userspace flows using the ovs-ofctl utility and also uses the
other_config:flow-restore-wait column to keep the traffic downtime to the
minimum. The ovs-ctl utility's "force-reload-kmod" function does all of the
above, but also replaces the old kernel module with the new one. Open vSwitch
startup scripts for Debian, XenServer and RHEL use ovs-ctl's functions and it
is recommended that these functions be used for other software platforms too.

Running the Testsuite
=====================

Open vSwitch includes a testsuite.  Before you submit patches
upstream, we advise that you run the tests and ensure that they pass.
If you add new features to Open vSwitch, then adding tests for those
features will ensure your features don't break as developers modify
other areas of Open vSwitch.

You must configure and build Open vSwitch (steps 1 through 3 in
"Building and Installing Open vSwitch for Linux, FreeBSD or NetBSD" above)
before you run the testsuite.  You do not need to install Open vSwitch
or to build or load the kernel module to run the testsuite.  You do
not need supervisor privilege to run the testsuite.

To run all the unit tests in Open vSwitch, one at a time:
      make check
This takes under 5 minutes on a modern desktop system.

To run all the unit tests in Open vSwitch, up to 8 in parallel:
      make check TESTSUITEFLAGS=-j8
This takes under a minute on a modern 4-core desktop system.

To see a list of all the available tests, run:
      make check TESTSUITEFLAGS=--list

To run only a subset of tests, e.g. test 123 and tests 477 through 484:
      make check TESTSUITEFLAGS='123 477-484'
(Tests do not have inter-dependencies, so you may run any subset.)

To run tests matching a keyword, e.g. "ovsdb":
      make check TESTSUITEFLAGS='-k ovsdb'

To see a complete list of test options:
      make check TESTSUITEFLAGS=--help

The results of a testing run are reported in tests/testsuite.log.
Please report test failures as bugs and include the testsuite.log in
your report.

If you have "valgrind" installed, then you can also run the testsuite
under valgrind by using "make check-valgrind" in place of "make
check".  All the same options are available via TESTSUITEFLAGS.  When
you do this, the "valgrind" results for test <N> are reported in files
named tests/testsuite.dir/<N>/valgrind.*.  You may find that the
valgrind results are easier to interpret if you put "-q" in
~/.valgrindrc, since that reduces the amount of output.

Sometimes a few tests may fail on some runs but not others.  This is
usually a bug in the testsuite, not a bug in Open vSwitch itself.  If
you find that a test fails intermittently, please report it, since the
developers may not have noticed.

Bug Reporting
-------------

Please report problems to bugs@openvswitch.org.