Software-Defined Networking tools for LXC (LinuX Containers)
Software-Defined Networking for Linux Containers
Pipework lets you connect together containers in arbitrarily complex scenarios. Pipework uses cgroups and namespace and works with "plain" LXC containers (created with
lxc-start), and with the awesome Docker.
Table of Contents generated with DocToc
If you use vCenter / VSphere / ESX / ESXi, set or ask your administrator to set Network Security Policies of the vSwitch as below:
After starting the guest OS and creating a bridge, you might also need to fine-tune the
br1interface as follows:
brctl stp br1 off(to disable the STP protocol and prevent the switch from disabling ports)
brctl setfd br1 2(to reduce the time taken by the
br1interface to go from blocking to forwarding state)
brctl setmaxage br1 0
If you use VirtualBox, you will have to update your VM network settings. Open the settings panel for the VM, go the the "Network" tab, pull down the "Advanced" settings. Here, the "Adapter Type" should be
pcnet(the full name is something like "PCnet-FAST III"), instead of the default
e1000(Intel PRO/1000). Also, "Promiscuous Mode" should be set to "Allow All".
If you don't do that, bridged containers won't work, because the virtual NIC will filter out all packets with a different MAC address. If you are running VirtualBox in headless mode, the command line equivalent of the above is
modifyvm --nicpromisc1 allow-all. If you are using Vagrant, you can add the following to the config for the same effect:
config.vm.provider "virtualbox" do |v| v.customize ['modifyvm', :id, '--nictype1', 'Am79C973'] v.customize ['modifyvm', :id, '--nicpromisc1', 'allow-all'] end
Note: it looks like some operating systems (e.g. CentOS 7) do not support
pcnetanymore. You might want to use the
virtio-net(Paravirtualized Network) interface with those.
Before using Pipework, please ask on the docker-user mailing list if there is a "native" way to achieve what you want to do without Pipework.
In the long run, Docker will allow complex scenarios, and Pipework should become obsolete.
If there is really no other way to plumb your containers together with the current version of Docker, then okay, let's see how we can help you!
The following examples show what Pipework can do for you and your containers.
Let's create two containers, running the web tier and the database tier:
APACHE=$(docker run -d apache /usr/sbin/httpd -D FOREGROUND) MYSQL=$(docker run -d mysql /usr/sbin/mysqld_safe)
Now, bring superpowers to the web tier:
pipework br1 $APACHE 192.168.1.1/24
br1in the docker host;
Now (drum roll), let's do this:
pipework br1 $MYSQL 192.168.1.2/24
br1, since it already exists;
Now, both containers can ping each other on the 192.168.1.0/24 subnet.
Pipework can resolve Docker containers names. If the container ID that you gave to Pipework cannot be found, Pipework will try to resolve it with
docker inspect. This makes it even simpler to use:
docker run -name web1 -d apache pipework br1 web1 192.168.12.23/24
Want to connect to those containers using their private addresses? Easy:
ip addr add 192.168.1.254/24 dev br1
By default pipework creates a new interface
eth1inside the container. In case you want to change this interface name like
eth2, e.g., to have more than one interface set by pipework, use:
pipework br1 -i eth2 ...
Note:: for InfiniBand IPoIB interfaces, the default interface name is
By default pipework will create a host-side interface with a fixed prefix but random suffix. If you would like to specify this interface name use the
-lflag (for local):
pipework br1 -i eth2 -l hostapp1 ...
The IP addresses given to
pipeworkare directly passed to the
ip addrtool; so you can append a subnet size using traditional CIDR notation.
pipework br1 $CONTAINERID 192.168.4.25/20
Don't forget that all containers should use the same subnet size; pipework is not clever enough to use your specified subnet size for the first container, and retain it to use it for the other containers.
If you want outbound traffic (i.e. when the containers connects to the outside world) to go through the interface managed by Pipework, you need to change the default route of the container.
This can be useful in some usecases, like traffic shaping, or if you want the container to use a specific outbound IP address.
This can be automated by Pipework, by adding the gateway address after the IP address and subnet mask:
pipework br1 $CONTAINERID 192.168.4.25/[email protected]
Let's pretend that you want to run two Hipache instances, listening on real interfaces eth2 and eth3, using specific (public) IP addresses. Easy!
pipework eth2 $(docker run -d hipache /usr/sbin/hipache) 184.108.40.206/24 pipework eth3 $(docker run -d hipache /usr/sbin/hipache) 220.127.116.11/24
Note that this will use
macvlansubinterfaces, so you can actually put multiple containers on the same physical interface. If you don't want to virtualize the interface, you can use the
--direct-physoption to namespace an interface exclusively to a container without using a macvlan bridge.
pipework --direct-phys eth1 $CONTAINERID 192.168.1.2/24
This is useful for assigning SR-IOV VFs to containers, but be aware of added latency when using the NIC to switch packets between containers on the same host.
In case you want to connect a local physical interface with a specific name inside the container, it will also rename the physical one, this behaviour is not idempotent:
pipework --direct-phys eth1 -i container0 $CONTAINERID 0/0 # second call would fail because physical interface eth1 has been renamed
We can use the interface MAC address to identify the interface the same way any time (udev networking rules use a similar method for interfaces persistent naming):
pipework --direct-phys mac:00:f3:15:4a:42:c8 -i container0 $CONTAINERID 0/0
If you use macvlan interfaces as shown in the previous paragraph, you will notice that the host will not be able to reach the containers over their macvlan interfaces. This is because traffic going in and out of macvlan interfaces is segregated from the "root" interface.
If you want to enable that kind of communication, no problem: just create a macvlan interface in your host, and move the IP address from the "normal" interface to the macvlan interface.
For instance, on a machine where
eth0is the main interface, and has address
10.1.1.123/24, with gateway
10.1.1.254, you would do this:
ip addr del 10.1.1.123/24 dev eth0 ip link add link eth0 dev eth0m type macvlan mode bridge ip link set eth0m up ip addr add 10.1.1.123/24 dev eth0m route add default gw 10.1.1.254
Then, you would start a container and assign it a macvlan interface the usual way:
CID=$(docker run -d ...) pipework eth0 $CID 10.1.1.234/[email protected]
Sometimes, you want the extra network interface to be up and running before starting your service. A dirty (and unreliable) solution would be to add a
sleepcommand before starting your service; but that could break in "interesting" ways if the server happens to be a bit slower at one point.
There is a better option: add the
pipeworkscript to your Docker image, and before starting the service, call
pipework --wait. It will wait until the
eth1interface is present and in
UPoperational state, then exit gracefully.
If you need to wait on an interface other than eth1, pass the -i flag like this:
pipework --wait -i ib0
If for some reason you want to set the IP address from within the container, you can use
0/0as the IP address. The interface will be created, connected to the network, and assigned to the container, but without configuring an IP address:
pipework br1 $CONTAINERID 0/0
If for some reason you want a dummy interface inside the container, you can add it like any other interface. Just set the host interface to the keyword dummy. All other options - IP, CIDR, gateway - function as normal.
pipework dummy $CONTAINERID 192.168.21.101/[email protected]
Of course, a gateway does not mean much in the context of a dummy interface, but there it is.
You can use DHCP to obtain the IP address of the new interface. Just specify the name of the DHCP client that you want to use instead on an IP address; for instance:
pipework eth1 $CONTAINERID dhclient
You can specify the following DHCP clients:
The first three are "normal" DHCP clients. They have to be installed on your host for this option to work. The last one works differently: it will run a DHCP client in a Docker container sharing its network namespace with your container. This allows to use DHCP configuration without worrying about installing the right DHCP client on your host. It will use the Docker
busyboximage and its embedded
The value of $CONTAINERID will be provided to the DHCP client to use as the hostname in the DHCP request. Depending on the configuration of your network's DHCP server, this may enable other machines on the network to access the container using the $CONTAINERID as a hostname; therefore, specifying $CONTAINERID as a container name rather than a container id may be more appropriate in this use-case.
You need three things for this to work correctly:
dhcpcp) must be installed on your Docker host (you don't have to install it in your containers, but it must be present on the host), unless you specify
dhcpas the client, in which case the Docker
busyboximage should be available;
The last item might be particularly relevant if you are trying to bridge your containers with a WPA-protected WiFi network. I'm not 100% sure about this, but I think that the WiFi access point will drop frames originating from unknown MAC addresses; meaning that you have to go through extra hoops if you want it to work properly.
It works fine on plain old wired Ethernet, though.
All of the DHCP options - udhcpc, dhcp, dhclient, dhcpcd - exit or are killed by pipework when they are done assigning a lease. This is to prevent zombie processes from existing after a container exits, but the dhcp client still exists.
However, if the container is long-running - longer than the life of the lease - then the lease will expire, no dhcp client renews the lease, and the container is stuck without a valid IP address.
To resolve this problem, you can cause the dhcp client to remain alive. The method depends on the dhcp client you use.
Note: If you use this option you will be responsible for finding and killing those dhcp client processes in the future. pipework is a one-time script; it is not intended to manage long-running processes for you.
In order to find the processes, you can look for pidfiles in the following locations:
$GUESTNAMEis the name or ID of the guest as you passed it to pipework on instantiation.
You can specify extra DHCP options to be passed to the DHCP client by adding them with a colon. For instance:
pipework eth1 $CONTAINERID dhcp:-f
This will tell Pipework to setup the interface using the DHCP client of the Docker
busyboximage, and pass
-fas an extra flag to this DHCP client. This flag instructs the client to remain in the foreground instead of going to the background. Let's see what this means.
Without this flag, a new container is started, in which the DHCP client is executed. The DHCP client obtains a lease, then goes to the background. When it goes to the background, the PID 1 in this container exits, causing the whole container to be terminated. As a result, the "pipeworked" container has its IP address, but the DHCP client has gone. On the up side, you don't have any cleanup to do; on the other, the DHCP lease will not be renewed, which could be problematic if you have short leases and the server and other clients don't validate their leases before using them.
With this flag, a new container is started, it runs the DHCP client just like before; but when it obtains the lease, it remains in the foreground. As a result, the lease will be properly renewed. However, when you terminate the "pipeworked" container, you should also take care of removing the container that runs the DHCP client. This can be seen as an advantage if you want to reuse this network stack even if the initial container is terminated.
If you need to specify the MAC address to be used (either by the
macvlansubinterface, or the
vethinterface), no problem. Just add it as the command-line, as the last argument:
pipework eth0 $(docker run -d haproxy) 192.168.1.2/24 26:2e:71:98:60:8f
This can be useful if your network environment requires whitelisting your hardware addresses (some hosting providers do that), or if you want to obtain a specific address from your DHCP server. Also, some projects like Orchestrator rely on static MAC-IPv6 bindings for DHCPv6:
pipework br0 $(docker run -d zerorpcworker) dhcp fa:de:b0:99:52:1c
Note: if you generate your own MAC addresses, try remember those two simple rules:
0, otherwise, you are defining a multicast address;
1, otherwise, you are using a globally unique (OUI enforced) address.
In other words, if your MAC address is
e. You can check Wikipedia if you want even more details.
If you want a consistent MAC address across container restarts, but don't want to have to keep track of the messy MAC addresses, ask pipework to generate an address for you based on a specified string, e.g. the hostname. This guarantees a consistent MAC address:
pipework eth0 dhcp U:
pipework will take some_string and hash it using MD5. It will then take the first 40 bits of the MD5 hash, add those to the locally administered prefix of 0x02, and create a unique MAC address.
For example, if your unique string is "myhost.foo.com", then the MAC address will always be
This is particularly useful in the case of DHCP, where you might want the container to stop and start, but always get the same address. Most DHCP servers will keep giving you a consistent IP address if the MAC address is consistent.
Note: Setting the MAC address of an IPoIB interface is not supported.
If you want to attach the container to a specific VLAN, the VLAN ID can be specified using the
[MAC]@VIDnotation in the MAC address parameter.
Note: VLAN attachment is currently only supported for containers to be attached to either an Open vSwitch bridge or a physical interface. Linux bridges are currently not supported.
The following will attach container zerorpcworker to the Open vSwitch bridge ovs0 and attach the container to VLAN ID 10.
pipework ovsbr0 $(docker run -d zerorpcworker) dhcp @10
If you want to add/delete/replace routes in the container, you can run any iproute2 route command via pipework.
All you have to do is set the interface to be
route, followed by the container ID or name, followed by the route command.
Here are some examples.
pipework route $CONTAINERID add 10.0.5.6/24 via 192.168.2.1 pipework route $CONTAINERID replace default via 10.2.3.5.78
Everything after the container ID (or name) will be run as an argument to
ip routeinside the container's namespace. Use the iproute2 man page.
If you want to add/delete/replace IP rules in the container, you can do the same thing with
ip rulethat you can with
Specify the interface to be
rule, followed by the container ID or name, followed by the rule command.
Here are some examples, to specify a route table:
pipework rule $CONTAINERID add from 172.19.0.2/32 table 1 pipework rule $CONTAINERID add to 172.19.0.2/32 table 1
Note that for these rules to work you first need to execute the following in your container:
echo "1 admin" >> /etc/iproute2/rt_tables
You can read more on using route tables, specifically to setup multiple NICs with different default gateways, here: https://kindlund.wordpress.com/2007/11/19/configuring-multiple-default-routes-in-linux/
If you want to use
tcfrom within the container namespace, you can do so with the command
pipework tc $CONTAINERID.
Example, to simulate 30% packet loss on
eth0within the container:
pipework tc $CONTAINERID qdisc add dev eth0 root netem loss 30%
If you want to attach a container to the Open vSwitch bridge, no problem.
ovs-vsctl list-br ovsbr0 pipework ovsbr0 $(docker run -d mysql /usr/sbin/mysqld_safe) 192.168.1.2/24
If the ovs bridge doesn't exist, it will be automatically created
Passing an IPoIB interface to a container is supported. The IPoIB device is created as a virtual device, similarly to how macvlan devices work. The interface also supports setting a partition key for the created virtual device.
The following will attach a container to ib0
pipework ib0 $CONTAINERID 10.10.10.10/24
The following will do the same but connect it to ib0 with pkey 0x8001
pipework ib0 $CONTAINERID 10.10.10.10/24 @8001
arpingis installed, it will be used to send a gratuitous ARP reply to the container's neighbors. This can be useful if the container doesn't emit any network traffic at all, and seems unreachable (but suddenly becomes reachable after it generates some traffic).
Note, however, that Ubuntu/Debian distributions contain two different
arpingpackages. The one you want is
When a container is terminated (the last process of the net namespace exits), the network interfaces are garbage collected. The interface in the container is automatically destroyed, and the interface in the docker host (part of the bridge) is then destroyed as well.
@dreamcat4 has built an amazing fork of pipework that can be integrated with other tools in the Docker ecosystem, like Compose or Crane. It can be used in "one shot," to create a bunch of network connections between containers; it can run in the background as a daemon, watching the Docker events API, and automatically invoke pipework when containers are started, and it can also expose pipework itself through an API.
For more info, check the dreamcat4/pipework image on the Docker Hub.
This README file is currently the only documentation for pipework. When updating it (specifically, when adding/removing/moving sections), please update the table of contents. This can be done very easily by just running:
This will build a container with
doctocand run it to regenerate the table of contents. That's it!