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
- Things to note
- LAMP stack with a private network between the MySQL and Apache containers
- Docker integration
- Peeking inside the private network
- Setting container internal interface
- Setting host interface name
- Using a different netmask
- Setting a default gateway
- Connect a container to a local physical interface
- Let the Docker host communicate over macvlan interfaces
- Wait for the network to be ready
- Add the interface without an IP address
- Add a dummy interface
- DHCP
- DHCP Options
- Specify a custom MAC address
- Virtual LAN (VLAN)
- Control routes
- Support Open vSwitch
- Support InfiniBand IPoIB
- Cleanup
- Integrating pipework with other tools
- About this file
If you use vCenter / VSphere / ESX / ESXi, set or ask your administrator to set Network Security Policies of the vSwitch as below:
- Promiscuous mode: Accept
- MAC address changes: Accept
- Forged transmits: Accept
After starting the guest OS and creating a bridge, you might also need to
fine-tune the br1
interface 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 thebr1
interface 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
pcnet
anymore. 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 $APACHE 192.168.1.1/24
This will:
- create a bridge named
br1
in the docker host; - add an interface named
eth1
to the$APACHE
container; - assign IP address 192.168.1.1 to this interface,
- connect said interface to
br1
.
Now (drum roll), let's do this:
pipework br1 $MYSQL 192.168.1.2/24
This will:
- not create a bridge named
br1
, since it already exists; - add an interface named
eth1
to the$MYSQL
container; - assign IP address 192.168.1.2 to this interface,
- connect said interface to
br1
.
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
VoilĂ !
By default pipework creates a new interface eth1
inside 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 ib0
and not eth1
.
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 -l
flag (for local):
pipework br1 -i eth2 -l hostapp1 ...
The IP addresses given to pipework
are directly passed to the ip addr
tool; so you can append a subnet size using traditional CIDR notation.
I.e.:
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 192.168.4.25/20@192.168.4.1
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) 50.19.169.157/24
pipework eth3 $(docker run -d hipache /usr/sbin/hipache) 107.22.140.5/24
Note that this will use macvlan
subinterfaces, 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-phys
option to namespace
an interface exclusively to a container without using a macvlan bridge.
pipework --direct-phys eth1 $CONTAINERID 192.168.1.2/24
This is useful for assigning SR-IOV VFs to containers, but be aware of added latency when using the NIC to switch packets between containers on the same host.
If you use macvlan interfaces as shown in the previous paragraph, you will notice that the host will not be able to reach the containers over their macvlan interfaces. This is because traffic going in and out of macvlan interfaces is segregated from the "root" interface.
If you want to enable that kind of communication, no problem: just create a macvlan interface in your host, and move the IP address from the "normal" interface to the macvlan interface.
For instance, on a machine where eth0
is the main interface, and has
address 10.1.1.123/24
, with gateway 10.1.1.254
, you would do this:
ip addr del 10.1.1.123/24 dev eth0
ip link add link eth0 dev eth0m type macvlan mode bridge
ip link set eth0m up
ip addr add 10.1.1.123/24 dev eth0m
route add default gw 10.1.1.254
Then, you would start a container and assign it a macvlan interface the usual way:
CID=$(docker run -d ...)
pipework eth0 $CID 10.1.1.234/24@10.1.1.254
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 sleep
command 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 pipework
script to your Docker image,
and before starting the service, call pipework --wait
. It will wait
until the eth1
interface is present and in UP
operational 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 0/0
as 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 192.168.21.101/24@192.168.21.1
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:
- dhclient
- udhcpc
- dhcpcd
- dhcp
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 busybox
image and its embedded udhcpc
client.
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:
- obviously, a DHCP server (in the example above, a DHCP server should
be listening on the network to which we are connected on
eth1
); - a DHCP client (either
udhcpc
,dhclient
ordhcpcp
) 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 specifydhcp
as the client, in which case the Dockerbusybox
image should be available; - the underlying network must support bridged frames.
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.
- dhcp: see the next section DHCP Options
- dhclient: use DHCP client
dhclient-f
- udhcpc: use DHCP client
udhcpc-f
- dhcpcd: not yet supported.
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:
- dhcp: see the next section DHCP Options
- dhclient: pidfiles in
/var/run/dhclient.$GUESTNAME.pid
- udhcpc: pidfiles in
/var/run/udhcpc.$GUESTNAME.pid
- dhcpcd: not yet supported
$GUESTNAME
is 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 busybox
image, and pass -f
as 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 macvlan
subinterface, or the veth
interface), 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:
- the lowest bit of the first byte should be
0
, otherwise, you are defining a multicast address; - the second lowest bit of the first byte should be
1
, otherwise, you are using a globally unique (OUI enforced) address.
In other words, if your MAC address is ?X:??:??:??:??:??
, X
should
be 2
, 6
, a
, or 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 <container> dhcp U:<some_string>
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 02:72:6c:cd:9b:8d
.
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]@VID
notation 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 route
inside the container's namespace. Use the iproute2 man page.
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
If arping
is 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 arping
packages. The one you want is iputils-arping
.
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:
docker-compose up
This will build a container with doctoc
and run it to regenerate the
table of contents. That's it!