1
0
Fork 0
mirror of https://github.com/moby/moby.git synced 2022-11-09 12:21:53 -05:00
moby--moby/docs/articles/networking.md
Mary Anthony 8fee1c2020 Enabled GitHub Flavored Markdown
GitHub flavored markdown is now supported for links and images. Also, ran LinkChecker and FileResolver. Yay!
Fixes from Spider check
Output for docker/docker now goes into engine directory

Signed-off-by: Mary Anthony <mary@docker.com>
2015-10-11 07:49:54 -07:00

50 KiB

Network configuration

Note: This document is outdated and needs a major overhaul.

Summary

When Docker starts, it creates a virtual interface named docker0 on the host machine. It randomly chooses an address and subnet from the private range defined by RFC 1918 that are not in use on the host machine, and assigns it to docker0. Docker made the choice 172.17.42.1/16 when I started it a few minutes ago, for example — a 16-bit netmask providing 65,534 addresses for the host machine and its containers. The MAC address is generated using the IP address allocated to the container to avoid ARP collisions, using a range from 02:42:ac:11:00:00 to 02:42:ac:11:ff:ff.

Note: This document discusses advanced networking configuration and options for Docker. In most cases you won't need this information. If you're looking to get started with a simpler explanation of Docker networking and an introduction to the concept of container linking see the Docker User Guide.

But docker0 is no ordinary interface. It is a virtual Ethernet bridge that automatically forwards packets between any other network interfaces that are attached to it. This lets containers communicate both with the host machine and with each other. Every time Docker creates a container, it creates a pair of “peer” interfaces that are like opposite ends of a pipe — a packet sent on one will be received on the other. It gives one of the peers to the container to become its eth0 interface and keeps the other peer, with a unique name like vethAQI2QT, out in the namespace of the host machine. By binding every veth* interface to the docker0 bridge, Docker creates a virtual subnet shared between the host machine and every Docker container.

The remaining sections of this document explain all of the ways that you can use Docker options and — in advanced cases — raw Linux networking commands to tweak, supplement, or entirely replace Docker's default networking configuration.

Quick guide to the options

Here is a quick list of the networking-related Docker command-line options, in case it helps you find the section below that you are looking for.

Some networking command-line options can only be supplied to the Docker server when it starts up, and cannot be changed once it is running:

There are three networking options that can be supplied either at startup or when docker run is invoked. When provided at startup, set the default value that docker run will later use if the options are not specified:

Finally, several networking options can only be provided when calling docker run because they specify something specific to one container:

To supply networking options to the Docker server at startup, use the DOCKER_OPTS variable in the Docker upstart configuration file. For Ubuntu, edit the variable in /etc/default/docker or /etc/sysconfig/docker for CentOS.

The following example illustrates how to configure Docker on Ubuntu to recognize a newly built bridge.

Edit the /etc/default/docker file:

$ echo 'DOCKER_OPTS="-b=bridge0"' >> /etc/default/docker 

Then restart the Docker server.

$ sudo service docker start

For additional information on bridges, see building your own bridge later on this page.

The following sections tackle all of the above topics in an order that we can move roughly from simplest to most complex.

Configuring DNS

How can Docker supply each container with a hostname and DNS configuration, without having to build a custom image with the hostname written inside? Its trick is to overlay three crucial /etc files inside the container with virtual files where it can write fresh information. You can see this by running mount inside a container:

$$ mount
...
/dev/disk/by-uuid/1fec...ebdf on /etc/hostname type ext4 ...
/dev/disk/by-uuid/1fec...ebdf on /etc/hosts type ext4 ...
/dev/disk/by-uuid/1fec...ebdf on /etc/resolv.conf type ext4 ...
...

This arrangement allows Docker to do clever things like keep resolv.conf up to date across all containers when the host machine receives new configuration over DHCP later. The exact details of how Docker maintains these files inside the container can change from one Docker version to the next, so you should leave the files themselves alone and use the following Docker options instead.

Four different options affect container domain name services.

  • -h HOSTNAME or --hostname=HOSTNAME — sets the hostname by which the container knows itself. This is written into /etc/hostname, into /etc/hosts as the name of the container's host-facing IP address, and is the name that /bin/bash inside the container will display inside its prompt. But the hostname is not easy to see from outside the container. It will not appear in docker ps nor in the /etc/hosts file of any other container.

  • --link=CONTAINER_NAME_or_ID:ALIAS — using this option as you run a container gives the new container's /etc/hosts an extra entry named ALIAS that points to the IP address of the container identified by CONTAINER_NAME_or_ID. This lets processes inside the new container connect to the hostname ALIAS without having to know its IP. The --link= option is discussed in more detail below, in the section Communication between containers. Because Docker may assign a different IP address to the linked containers on restart, Docker updates the ALIAS entry in the /etc/hosts file of the recipient containers.

  • --dns=IP_ADDRESS... — sets the IP addresses added as server lines to the container's /etc/resolv.conf file. Processes in the container, when confronted with a hostname not in /etc/hosts, will connect to these IP addresses on port 53 looking for name resolution services.

  • --dns-search=DOMAIN... — sets the domain names that are searched when a bare unqualified hostname is used inside of the container, by writing search lines into the container's /etc/resolv.conf. When a container process attempts to access host and the search domain example.com is set, for instance, the DNS logic will not only look up host but also host.example.com. Use --dns-search=. if you don't wish to set the search domain.

  • --dns-opt=OPTION... — sets the options used by DNS resolvers by writing an options line into the container's /etc/resolv.conf. See documentation for resolv.conf for a list of valid options.

Regarding DNS settings, in the absence of the --dns=IP_ADDRESS..., --dns-search=DOMAIN..., or --dns-opt=OPTION... options, Docker makes each container's /etc/resolv.conf look like the /etc/resolv.conf of the host machine (where the docker daemon runs). When creating the container's /etc/resolv.conf, the daemon filters out all localhost IP address nameserver entries from the host's original file.

Filtering is necessary because all localhost addresses on the host are unreachable from the container's network. After this filtering, if there are no more nameserver entries left in the container's /etc/resolv.conf file, the daemon adds public Google DNS nameservers (8.8.8.8 and 8.8.4.4) to the container's DNS configuration. If IPv6 is enabled on the daemon, the public IPv6 Google DNS nameservers will also be added (2001:4860:4860::8888 and 2001:4860:4860::8844).

Note

: If you need access to a host's localhost resolver, you must modify your DNS service on the host to listen on a non-localhost address that is reachable from within the container.

You might wonder what happens when the host machine's /etc/resolv.conf file changes. The docker daemon has a file change notifier active which will watch for changes to the host DNS configuration.

Note

: The file change notifier relies on the Linux kernel's inotify feature. Because this feature is currently incompatible with the overlay filesystem driver, a Docker daemon using "overlay" will not be able to take advantage of the /etc/resolv.conf auto-update feature.

When the host file changes, all stopped containers which have a matching resolv.conf to the host will be updated immediately to this newest host configuration. Containers which are running when the host configuration changes will need to stop and start to pick up the host changes due to lack of a facility to ensure atomic writes of the resolv.conf file while the container is running. If the container's resolv.conf has been edited since it was started with the default configuration, no replacement will be attempted as it would overwrite the changes performed by the container. If the options (--dns, --dns-search, or --dns-opt) have been used to modify the default host configuration, then the replacement with an updated host's /etc/resolv.conf will not happen as well.

Note

: For containers which were created prior to the implementation of the /etc/resolv.conf update feature in Docker 1.5.0: those containers will not receive updates when the host resolv.conf file changes. Only containers created with Docker 1.5.0 and above will utilize this auto-update feature.

Communication between containers and the wider world

Whether a container can talk to the world is governed by two factors.

  1. Is the host machine willing to forward IP packets? This is governed by the ip_forward system parameter. Packets can only pass between containers if this parameter is 1. Usually you will simply leave the Docker server at its default setting --ip-forward=true and Docker will go set ip_forward to 1 for you when the server starts up. If you set --ip-forward=false and your system's kernel has it enabled, the --ip-forward=false option has no effect. To check the setting on your kernel or to turn it on manually:

    $ sysctl net.ipv4.conf.all.forwarding
    net.ipv4.conf.all.forwarding = 0
    $ sysctl net.ipv4.conf.all.forwarding=1
    $ sysctl net.ipv4.conf.all.forwarding
    net.ipv4.conf.all.forwarding = 1
    

    Many using Docker will want ip_forward to be on, to at least make communication possible between containers and the wider world.

    May also be needed for inter-container communication if you are in a multiple bridge setup.

  2. Do your iptables allow this particular connection? Docker will never make changes to your system iptables rules if you set --iptables=false when the daemon starts. Otherwise the Docker server will append forwarding rules to the DOCKER filter chain.

Docker will not delete or modify any pre-existing rules from the DOCKER filter chain. This allows the user to create in advance any rules required to further restrict access to the containers.

Docker's forward rules permit all external source IPs by default. To allow only a specific IP or network to access the containers, insert a negated rule at the top of the DOCKER filter chain. For example, to restrict external access such that only source IP 8.8.8.8 can access the containers, the following rule could be added:

$ iptables -I DOCKER -i ext_if ! -s 8.8.8.8 -j DROP

Communication between containers

Whether two containers can communicate is governed, at the operating system level, by two factors.

  1. Does the network topology even connect the containers' network interfaces? By default Docker will attach all containers to a single docker0 bridge, providing a path for packets to travel between them. See the later sections of this document for other possible topologies.

  2. Do your iptables allow this particular connection? Docker will never make changes to your system iptables rules if you set --iptables=false when the daemon starts. Otherwise the Docker server will add a default rule to the FORWARD chain with a blanket ACCEPT policy if you retain the default --icc=true, or else will set the policy to DROP if --icc=false.

It is a strategic question whether to leave --icc=true or change it to --icc=false so that iptables will protect other containers — and the main host — from having arbitrary ports probed or accessed by a container that gets compromised.

If you choose the most secure setting of --icc=false, then how can containers communicate in those cases where you want them to provide each other services?

The answer is the --link=CONTAINER_NAME_or_ID:ALIAS option, which was mentioned in the previous section because of its effect upon name services. If the Docker daemon is running with both --icc=false and --iptables=true then, when it sees docker run invoked with the --link= option, the Docker server will insert a pair of iptables ACCEPT rules so that the new container can connect to the ports exposed by the other container — the ports that it mentioned in the EXPOSE lines of its Dockerfile. Docker has more documentation on this subject — see the linking Docker containers page for further details.

Note

: The value CONTAINER_NAME in --link= must either be an auto-assigned Docker name like stupefied_pare or else the name you assigned with --name= when you ran docker run. It cannot be a hostname, which Docker will not recognize in the context of the --link= option.

You can run the iptables command on your Docker host to see whether the FORWARD chain has a default policy of ACCEPT or DROP:

# When --icc=false, you should see a DROP rule:

$ sudo iptables -L -n
...
Chain FORWARD (policy ACCEPT)
target     prot opt source               destination
DOCKER     all  --  0.0.0.0/0            0.0.0.0/0
DROP       all  --  0.0.0.0/0            0.0.0.0/0
...

# When a --link= has been created under --icc=false,
# you should see port-specific ACCEPT rules overriding
# the subsequent DROP policy for all other packets:

$ sudo iptables -L -n
...
Chain FORWARD (policy ACCEPT)
target     prot opt source               destination
DOCKER     all  --  0.0.0.0/0            0.0.0.0/0
DROP       all  --  0.0.0.0/0            0.0.0.0/0

Chain DOCKER (1 references)
target     prot opt source               destination
ACCEPT     tcp  --  172.17.0.2           172.17.0.3           tcp spt:80
ACCEPT     tcp  --  172.17.0.3           172.17.0.2           tcp dpt:80

Note

: Docker is careful that its host-wide iptables rules fully expose containers to each other's raw IP addresses, so connections from one container to another should always appear to be originating from the first container's own IP address.

Binding container ports to the host

By default Docker containers can make connections to the outside world, but the outside world cannot connect to containers. Each outgoing connection will appear to originate from one of the host machine's own IP addresses thanks to an iptables masquerading rule on the host machine that the Docker server creates when it starts:

# You can see that the Docker server creates a
# masquerade rule that let containers connect
# to IP addresses in the outside world:

$ sudo iptables -t nat -L -n
...
Chain POSTROUTING (policy ACCEPT)
target     prot opt source               destination
MASQUERADE  all  --  172.17.0.0/16       0.0.0.0/0
...

But if you want containers to accept incoming connections, you will need to provide special options when invoking docker run. These options are covered in more detail in the Docker User Guide page. There are two approaches.

First, you can supply -P or --publish-all=true|false to docker run which is a blanket operation that identifies every port with an EXPOSE line in the image's Dockerfile or --expose <port> commandline flag and maps it to a host port somewhere within an ephemeral port range. The docker port command then needs to be used to inspect created mapping. The ephemeral port range is configured by /proc/sys/net/ipv4/ip_local_port_range kernel parameter, typically ranging from 32768 to 61000.

Mapping can be specified explicitly using -p SPEC or --publish=SPEC option. It allows you to particularize which port on docker server - which can be any port at all, not just one within the ephemeral port range — you want mapped to which port in the container.

Either way, you should be able to peek at what Docker has accomplished in your network stack by examining your NAT tables.

# What your NAT rules might look like when Docker
# is finished setting up a -P forward:

$ iptables -t nat -L -n
...
Chain DOCKER (2 references)
target     prot opt source               destination
DNAT       tcp  --  0.0.0.0/0            0.0.0.0/0            tcp dpt:49153 to:172.17.0.2:80

# What your NAT rules might look like when Docker
# is finished setting up a -p 80:80 forward:

Chain DOCKER (2 references)
target     prot opt source               destination
DNAT       tcp  --  0.0.0.0/0            0.0.0.0/0            tcp dpt:80 to:172.17.0.2:80

You can see that Docker has exposed these container ports on 0.0.0.0, the wildcard IP address that will match any possible incoming port on the host machine. If you want to be more restrictive and only allow container services to be contacted through a specific external interface on the host machine, you have two choices. When you invoke docker run you can use either -p IP:host_port:container_port or -p IP::port to specify the external interface for one particular binding.

Or if you always want Docker port forwards to bind to one specific IP address, you can edit your system-wide Docker server settings and add the option --ip=IP_ADDRESS. Remember to restart your Docker server after editing this setting.

Note

: With hairpin NAT enabled (--userland-proxy=false), containers port exposure is achieved purely through iptables rules, and no attempt to bind the exposed port is ever made. This means that nothing prevents shadowing a previously listening service outside of Docker through exposing the same port for a container. In such conflicting situation, Docker created iptables rules will take precedence and route to the container.

The --userland-proxy parameter, true by default, provides a userland implementation for inter-container and outside-to-container communication. When disabled, Docker uses both an additional MASQUERADE iptable rule and the net.ipv4.route_localnet kernel parameter which allow the host machine to connect to a local container exposed port through the commonly used loopback address: this alternative is preferred for performance reasons.

Again, this topic is covered without all of these low-level networking details in the Docker User Guide document if you would like to use that as your port redirection reference instead.

IPv6

As we are running out of IPv4 addresses the IETF has standardized an IPv4 successor, Internet Protocol Version 6 , in RFC 2460. Both protocols, IPv4 and IPv6, reside on layer 3 of the OSI model.

IPv6 with Docker

By default, the Docker server configures the container network for IPv4 only. You can enable IPv4/IPv6 dualstack support by running the Docker daemon with the --ipv6 flag. Docker will set up the bridge docker0 with the IPv6 link-local address fe80::1.

By default, containers that are created will only get a link-local IPv6 address. To assign globally routable IPv6 addresses to your containers you have to specify an IPv6 subnet to pick the addresses from. Set the IPv6 subnet via the --fixed-cidr-v6 parameter when starting Docker daemon:

docker daemon --ipv6 --fixed-cidr-v6="2001:db8:1::/64"

The subnet for Docker containers should at least have a size of /80. This way an IPv6 address can end with the container's MAC address and you prevent NDP neighbor cache invalidation issues in the Docker layer.

With the --fixed-cidr-v6 parameter set Docker will add a new route to the routing table. Further IPv6 routing will be enabled (you may prevent this by starting Docker daemon with --ip-forward=false):

$ ip -6 route add 2001:db8:1::/64 dev docker0
$ sysctl net.ipv6.conf.default.forwarding=1
$ sysctl net.ipv6.conf.all.forwarding=1

All traffic to the subnet 2001:db8:1::/64 will now be routed via the docker0 interface.

Be aware that IPv6 forwarding may interfere with your existing IPv6 configuration: If you are using Router Advertisements to get IPv6 settings for your host's interfaces you should set accept_ra to 2. Otherwise IPv6 enabled forwarding will result in rejecting Router Advertisements. E.g., if you want to configure eth0 via Router Advertisements you should set:

$ sysctl net.ipv6.conf.eth0.accept_ra=2

Every new container will get an IPv6 address from the defined subnet. Further a default route will be added on eth0 in the container via the address specified by the daemon option --default-gateway-v6 if present, otherwise via fe80::1:

docker run -it ubuntu bash -c "ip -6 addr show dev eth0; ip -6 route show"

15: eth0: <BROADCAST,UP,LOWER_UP> mtu 1500
   inet6 2001:db8:1:0:0:242:ac11:3/64 scope global
      valid_lft forever preferred_lft forever
   inet6 fe80::42:acff:fe11:3/64 scope link
      valid_lft forever preferred_lft forever

2001:db8:1::/64 dev eth0  proto kernel  metric 256
fe80::/64 dev eth0  proto kernel  metric 256
default via fe80::1 dev eth0  metric 1024

In this example the Docker container is assigned a link-local address with the network suffix /64 (here: fe80::42:acff:fe11:3/64) and a globally routable IPv6 address (here: 2001:db8:1:0:0:242:ac11:3/64). The container will create connections to addresses outside of the 2001:db8:1::/64 network via the link-local gateway at fe80::1 on eth0.

Often servers or virtual machines get a /64 IPv6 subnet assigned (e.g. 2001:db8:23:42::/64). In this case you can split it up further and provide Docker a /80 subnet while using a separate /80 subnet for other applications on the host:

In this setup the subnet 2001:db8:23:42::/80 with a range from 2001:db8:23:42:0:0:0:0 to 2001:db8:23:42:0:ffff:ffff:ffff is attached to eth0, with the host listening at 2001:db8:23:42::1. The subnet 2001:db8:23:42:1::/80 with an address range from 2001:db8:23:42:1:0:0:0 to 2001:db8:23:42:1:ffff:ffff:ffff is attached to docker0 and will be used by containers.

Using NDP proxying

If your Docker host is only part of an IPv6 subnet but has not got an IPv6 subnet assigned you can use NDP proxying to connect your containers via IPv6 to the internet. For example your host has the IPv6 address 2001:db8::c001, is part of the subnet 2001:db8::/64 and your IaaS provider allows you to configure the IPv6 addresses 2001:db8::c000 to 2001:db8::c00f:

$ ip -6 addr show
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536
    inet6 ::1/128 scope host
       valid_lft forever preferred_lft forever
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qlen 1000
    inet6 2001:db8::c001/64 scope global
       valid_lft forever preferred_lft forever
    inet6 fe80::601:3fff:fea1:9c01/64 scope link
       valid_lft forever preferred_lft forever

Let's split up the configurable address range into two subnets 2001:db8::c000/125 and 2001:db8::c008/125. The first one can be used by the host itself, the latter by Docker:

docker daemon --ipv6 --fixed-cidr-v6 2001:db8::c008/125

You notice the Docker subnet is within the subnet managed by your router that is connected to eth0. This means all devices (containers) with the addresses from the Docker subnet are expected to be found within the router subnet. Therefore the router thinks it can talk to these containers directly.

As soon as the router wants to send an IPv6 packet to the first container it will transmit a neighbor solicitation request, asking, who has 2001:db8::c009? But it will get no answer because no one on this subnet has this address. The container with this address is hidden behind the Docker host. The Docker host has to listen to neighbor solicitation requests for the container address and send a response that itself is the device that is responsible for the address. This is done by a Kernel feature called NDP Proxy. You can enable it by executing

$ sysctl net.ipv6.conf.eth0.proxy_ndp=1

Now you can add the container's IPv6 address to the NDP proxy table:

$ ip -6 neigh add proxy 2001:db8::c009 dev eth0

This command tells the Kernel to answer to incoming neighbor solicitation requests regarding the IPv6 address 2001:db8::c009 on the device eth0. As a consequence of this all traffic to this IPv6 address will go into the Docker host and it will forward it according to its routing table via the docker0 device to the container network:

$ ip -6 route show
2001:db8::c008/125 dev docker0  metric 1
2001:db8::/64 dev eth0  proto kernel  metric 256

You have to execute the ip -6 neigh add proxy ... command for every IPv6 address in your Docker subnet. Unfortunately there is no functionality for adding a whole subnet by executing one command. An alternative approach would be to use an NDP proxy daemon such as ndppd.

Docker IPv6 cluster

Switched network environment

Using routable IPv6 addresses allows you to realize communication between containers on different hosts. Let's have a look at a simple Docker IPv6 cluster example:

The Docker hosts are in the 2001:db8:0::/64 subnet. Host1 is configured to provide addresses from the 2001:db8:1::/64 subnet to its containers. It has three routes configured:

  • Route all traffic to 2001:db8:0::/64 via eth0
  • Route all traffic to 2001:db8:1::/64 via docker0
  • Route all traffic to 2001:db8:2::/64 via Host2 with IP 2001:db8::2

Host1 also acts as a router on OSI layer 3. When one of the network clients tries to contact a target that is specified in Host1's routing table Host1 will forward the traffic accordingly. It acts as a router for all networks it knows: 2001:db8::/64, 2001:db8:1::/64 and 2001:db8:2::/64.

On Host2 we have nearly the same configuration. Host2's containers will get IPv6 addresses from 2001:db8:2::/64. Host2 has three routes configured:

  • Route all traffic to 2001:db8:0::/64 via eth0
  • Route all traffic to 2001:db8:2::/64 via docker0
  • Route all traffic to 2001:db8:1::/64 via Host1 with IP 2001:db8:0::1

The difference to Host1 is that the network 2001:db8:2::/64 is directly attached to the host via its docker0 interface whereas it reaches 2001:db8:1::/64 via Host1's IPv6 address 2001:db8::1.

This way every container is able to contact every other container. The containers Container1-* share the same subnet and contact each other directly. The traffic between Container1-* and Container2-* will be routed via Host1 and Host2 because those containers do not share the same subnet.

In a switched environment every host has to know all routes to every subnet. You always have to update the hosts' routing tables once you add or remove a host to the cluster.

Every configuration in the diagram that is shown below the dashed line is handled by Docker: The docker0 bridge IP address configuration, the route to the Docker subnet on the host, the container IP addresses and the routes on the containers. The configuration above the line is up to the user and can be adapted to the individual environment.

Routed network environment

In a routed network environment you replace the layer 2 switch with a layer 3 router. Now the hosts just have to know their default gateway (the router) and the route to their own containers (managed by Docker). The router holds all routing information about the Docker subnets. When you add or remove a host to this environment you just have to update the routing table in the router - not on every host.

In this scenario containers of the same host can communicate directly with each other. The traffic between containers on different hosts will be routed via their hosts and the router. For example packet from Container1-1 to Container2-1 will be routed through Host1, Router and Host2 until it arrives at Container2-1.

To keep the IPv6 addresses short in this example a /48 network is assigned to every host. The hosts use a /64 subnet of this for its own services and one for Docker. When adding a third host you would add a route for the subnet 2001:db8:3::/48 in the router and configure Docker on Host3 with --fixed-cidr-v6=2001:db8:3:1::/64.

Remember the subnet for Docker containers should at least have a size of /80. This way an IPv6 address can end with the container's MAC address and you prevent NDP neighbor cache invalidation issues in the Docker layer. So if you have a /64 for your whole environment use /78 subnets for the hosts and /80 for the containers. This way you can use 4096 hosts with 16 /80 subnets each.

Every configuration in the diagram that is visualized below the dashed line is handled by Docker: The docker0 bridge IP address configuration, the route to the Docker subnet on the host, the container IP addresses and the routes on the containers. The configuration above the line is up to the user and can be adapted to the individual environment.

Customizing docker0

By default, the Docker server creates and configures the host system's docker0 interface as an Ethernet bridge inside the Linux kernel that can pass packets back and forth between other physical or virtual network interfaces so that they behave as a single Ethernet network.

Docker configures docker0 with an IP address, netmask and IP allocation range. The host machine can both receive and send packets to containers connected to the bridge, and gives it an MTU — the maximum transmission unit or largest packet length that the interface will allow — of either 1,500 bytes or else a more specific value copied from the Docker host's interface that supports its default route. These options are configurable at server startup:

  • --bip=CIDR — supply a specific IP address and netmask for the docker0 bridge, using standard CIDR notation like 192.168.1.5/24.

  • --fixed-cidr=CIDR — restrict the IP range from the docker0 subnet, using the standard CIDR notation like 172.167.1.0/28. This range must be an IPv4 range for fixed IPs (ex: 10.20.0.0/16) and must be a subset of the bridge IP range (docker0 or set using --bridge). For example with --fixed-cidr=192.168.1.0/25, IPs for your containers will be chosen from the first half of 192.168.1.0/24 subnet.

  • --mtu=BYTES — override the maximum packet length on docker0.

Once you have one or more containers up and running, you can confirm that Docker has properly connected them to the docker0 bridge by running the brctl command on the host machine and looking at the interfaces column of the output. Here is a host with two different containers connected:

# Display bridge info

$ sudo brctl show
bridge name     bridge id               STP enabled     interfaces
docker0         8000.3a1d7362b4ee       no              veth65f9
                                                        vethdda6

If the brctl command is not installed on your Docker host, then on Ubuntu you should be able to run sudo apt-get install bridge-utils to install it.

Finally, the docker0 Ethernet bridge settings are used every time you create a new container. Docker selects a free IP address from the range available on the bridge each time you docker run a new container, and configures the container's eth0 interface with that IP address and the bridge's netmask. The Docker host's own IP address on the bridge is used as the default gateway by which each container reaches the rest of the Internet.

# The network, as seen from a container

$ docker run -i -t --rm base /bin/bash

$$ ip addr show eth0
24: eth0: <BROADCAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast state UP group default qlen 1000
    link/ether 32:6f:e0:35:57:91 brd ff:ff:ff:ff:ff:ff
    inet 172.17.0.3/16 scope global eth0
       valid_lft forever preferred_lft forever
    inet6 fe80::306f:e0ff:fe35:5791/64 scope link
       valid_lft forever preferred_lft forever

$$ ip route
default via 172.17.42.1 dev eth0
172.17.0.0/16 dev eth0  proto kernel  scope link  src 172.17.0.3

$$ exit

Remember that the Docker host will not be willing to forward container packets out on to the Internet unless its ip_forward system setting is 1 — see the section above on Communication between containers for details.

Building your own bridge

If you want to take Docker out of the business of creating its own Ethernet bridge entirely, you can set up your own bridge before starting Docker and use -b BRIDGE or --bridge=BRIDGE to tell Docker to use your bridge instead. If you already have Docker up and running with its old docker0 still configured, you will probably want to begin by stopping the service and removing the interface:

# Stopping Docker and removing docker0

$ sudo service docker stop
$ sudo ip link set dev docker0 down
$ sudo brctl delbr docker0
$ sudo iptables -t nat -F POSTROUTING

Then, before starting the Docker service, create your own bridge and give it whatever configuration you want. Here we will create a simple enough bridge that we really could just have used the options in the previous section to customize docker0, but it will be enough to illustrate the technique.

# Create our own bridge

$ sudo brctl addbr bridge0
$ sudo ip addr add 192.168.5.1/24 dev bridge0
$ sudo ip link set dev bridge0 up

# Confirming that our bridge is up and running

$ ip addr show bridge0
4: bridge0: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state UP group default
    link/ether 66:38:d0:0d:76:18 brd ff:ff:ff:ff:ff:ff
    inet 192.168.5.1/24 scope global bridge0
       valid_lft forever preferred_lft forever

# Tell Docker about it and restart (on Ubuntu)

$ echo 'DOCKER_OPTS="-b=bridge0"' >> /etc/default/docker
$ sudo service docker start

# Confirming new outgoing NAT masquerade is set up

$ sudo iptables -t nat -L -n
...
Chain POSTROUTING (policy ACCEPT)
target     prot opt source               destination
MASQUERADE  all  --  192.168.5.0/24      0.0.0.0/0

The result should be that the Docker server starts successfully and is now prepared to bind containers to the new bridge. After pausing to verify the bridge's configuration, try creating a container — you will see that its IP address is in your new IP address range, which Docker will have auto-detected.

Just as we learned in the previous section, you can use the brctl show command to see Docker add and remove interfaces from the bridge as you start and stop containers, and can run ip addr and ip route inside a container to see that it has been given an address in the bridge's IP address range and has been told to use the Docker host's IP address on the bridge as its default gateway to the rest of the Internet.

How Docker networks a container

While Docker is under active development and continues to tweak and improve its network configuration logic, the shell commands in this section are rough equivalents to the steps that Docker takes when configuring networking for each new container.

Let's review a few basics.

To communicate using the Internet Protocol (IP), a machine needs access to at least one network interface at which packets can be sent and received, and a routing table that defines the range of IP addresses reachable through that interface. Network interfaces do not have to be physical devices. In fact, the lo loopback interface available on every Linux machine (and inside each Docker container) is entirely virtual — the Linux kernel simply copies loopback packets directly from the sender's memory into the receiver's memory.

Docker uses special virtual interfaces to let containers communicate with the host machine — pairs of virtual interfaces called “peers” that are linked inside of the host machine's kernel so that packets can travel between them. They are simple to create, as we will see in a moment.

The steps with which Docker configures a container are:

  1. Create a pair of peer virtual interfaces.

  2. Give one of them a unique name like veth65f9, keep it inside of the main Docker host, and bind it to docker0 or whatever bridge Docker is supposed to be using.

  3. Toss the other interface over the wall into the new container (which will already have been provided with an lo interface) and rename it to the much prettier name eth0 since, inside of the container's separate and unique network interface namespace, there are no physical interfaces with which this name could collide.

  4. Set the interface's MAC address according to the --mac-address parameter or generate a random one.

  5. Give the container's eth0 a new IP address from within the bridge's range of network addresses. The default route is set to the IP address passed to the Docker daemon using the --default-gateway option if specified, otherwise to the IP address that the Docker host owns on the bridge. The MAC address is generated from the IP address unless otherwise specified. This prevents ARP cache invalidation problems, when a new container comes up with an IP used in the past by another container with another MAC.

With these steps complete, the container now possesses an eth0 (virtual) network card and will find itself able to communicate with other containers and the rest of the Internet.

You can opt out of the above process for a particular container by giving the --net= option to docker run, which takes four possible values.

  • --net=bridge — The default action, that connects the container to the Docker bridge as described above.

  • --net=host — Tells Docker to skip placing the container inside of a separate network stack. In essence, this choice tells Docker to not containerize the container's networking! While container processes will still be confined to their own filesystem and process list and resource limits, a quick ip addr command will show you that, network-wise, they live “outside” in the main Docker host and have full access to its network interfaces. Note that this does not let the container reconfigure the host network stack — that would require --privileged=true — but it does let container processes open low-numbered ports like any other root process. It also allows the container to access local network services like D-bus. This can lead to processes in the container being able to do unexpected things like restart your computer. You should use this option with caution.

  • --net=container:NAME_or_ID — Tells Docker to put this container's processes inside of the network stack that has already been created inside of another container. The new container's processes will be confined to their own filesystem and process list and resource limits, but will share the same IP address and port numbers as the first container, and processes on the two containers will be able to connect to each other over the loopback interface.

  • --net=none — Tells Docker to put the container inside of its own network stack but not to take any steps to configure its network, leaving you free to build any of the custom configurations explored in the last few sections of this document.

To get an idea of the steps that are necessary if you use --net=none as described in that last bullet point, here are the commands that you would run to reach roughly the same configuration as if you had let Docker do all of the configuration:

# At one shell, start a container and
# leave its shell idle and running

$ docker run -i -t --rm --net=none base /bin/bash
root@63f36fc01b5f:/#

# At another shell, learn the container process ID
# and create its namespace entry in /var/run/netns/
# for the "ip netns" command we will be using below

$ docker inspect -f '{{.State.Pid}}' 63f36fc01b5f
2778
$ pid=2778
$ sudo mkdir -p /var/run/netns
$ sudo ln -s /proc/$pid/ns/net /var/run/netns/$pid

# Check the bridge's IP address and netmask

$ ip addr show docker0
21: docker0: ...
inet 172.17.42.1/16 scope global docker0
...

# Create a pair of "peer" interfaces A and B,
# bind the A end to the bridge, and bring it up

$ sudo ip link add A type veth peer name B
$ sudo brctl addif docker0 A
$ sudo ip link set A up

# Place B inside the container's network namespace,
# rename to eth0, and activate it with a free IP

$ sudo ip link set B netns $pid
$ sudo ip netns exec $pid ip link set dev B name eth0
$ sudo ip netns exec $pid ip link set eth0 address 12:34:56:78:9a:bc
$ sudo ip netns exec $pid ip link set eth0 up
$ sudo ip netns exec $pid ip addr add 172.17.42.99/16 dev eth0
$ sudo ip netns exec $pid ip route add default via 172.17.42.1

At this point your container should be able to perform networking operations as usual.

When you finally exit the shell and Docker cleans up the container, the network namespace is destroyed along with our virtual eth0 — whose destruction in turn destroys interface A out in the Docker host and automatically un-registers it from the docker0 bridge. So everything gets cleaned up without our having to run any extra commands! Well, almost everything:

# Clean up dangling symlinks in /var/run/netns

find -L /var/run/netns -type l -delete

Also note that while the script above used modern ip command instead of old deprecated wrappers like ipconfig and route, these older commands would also have worked inside of our container. The ip addr command can be typed as ip a if you are in a hurry.

Finally, note the importance of the ip netns exec command, which let us reach inside and configure a network namespace as root. The same commands would not have worked if run inside of the container, because part of safe containerization is that Docker strips container processes of the right to configure their own networks. Using ip netns exec is what let us finish up the configuration without having to take the dangerous step of running the container itself with --privileged=true.

Tools and examples

Before diving into the following sections on custom network topologies, you might be interested in glancing at a few external tools or examples of the same kinds of configuration. Here are two:

  • Jérôme Petazzoni has created a pipework shell script to help you connect together containers in arbitrarily complex scenarios: https://github.com/jpetazzo/pipework

  • Brandon Rhodes has created a whole network topology of Docker containers for the next edition of Foundations of Python Network Programming that includes routing, NAT'd firewalls, and servers that offer HTTP, SMTP, POP, IMAP, Telnet, SSH, and FTP: https://github.com/brandon-rhodes/fopnp/tree/m/playground

Both tools use networking commands very much like the ones you saw in the previous section, and will see in the following sections.

Building a point-to-point connection

By default, Docker attaches all containers to the virtual subnet implemented by docker0. You can create containers that are each connected to some different virtual subnet by creating your own bridge as shown in Building your own bridge, starting each container with docker run --net=none, and then attaching the containers to your bridge with the shell commands shown in How Docker networks a container.

But sometimes you want two particular containers to be able to communicate directly without the added complexity of both being bound to a host-wide Ethernet bridge.

The solution is simple: when you create your pair of peer interfaces, simply throw both of them into containers, and configure them as classic point-to-point links. The two containers will then be able to communicate directly (provided you manage to tell each container the other's IP address, of course). You might adjust the instructions of the previous section to go something like this:

# Start up two containers in two terminal windows

$ docker run -i -t --rm --net=none base /bin/bash
root@1f1f4c1f931a:/#

$ docker run -i -t --rm --net=none base /bin/bash
root@12e343489d2f:/#

# Learn the container process IDs
# and create their namespace entries

$ docker inspect -f '{{.State.Pid}}' 1f1f4c1f931a
2989
$ docker inspect -f '{{.State.Pid}}' 12e343489d2f
3004
$ sudo mkdir -p /var/run/netns
$ sudo ln -s /proc/2989/ns/net /var/run/netns/2989
$ sudo ln -s /proc/3004/ns/net /var/run/netns/3004

# Create the "peer" interfaces and hand them out

$ sudo ip link add A type veth peer name B

$ sudo ip link set A netns 2989
$ sudo ip netns exec 2989 ip addr add 10.1.1.1/32 dev A
$ sudo ip netns exec 2989 ip link set A up
$ sudo ip netns exec 2989 ip route add 10.1.1.2/32 dev A

$ sudo ip link set B netns 3004
$ sudo ip netns exec 3004 ip addr add 10.1.1.2/32 dev B
$ sudo ip netns exec 3004 ip link set B up
$ sudo ip netns exec 3004 ip route add 10.1.1.1/32 dev B

The two containers should now be able to ping each other and make connections successfully. Point-to-point links like this do not depend on a subnet nor a netmask, but on the bare assertion made by ip route that some other single IP address is connected to a particular network interface.

Note that point-to-point links can be safely combined with other kinds of network connectivity — there is no need to start the containers with --net=none if you want point-to-point links to be an addition to the container's normal networking instead of a replacement.

A final permutation of this pattern is to create the point-to-point link between the Docker host and one container, which would allow the host to communicate with that one container on some single IP address and thus communicate “out-of-band” of the bridge that connects the other, more usual containers. But unless you have very specific networking needs that drive you to such a solution, it is probably far preferable to use --icc=false to lock down inter-container communication, as we explored earlier.

Editing networking config files

Starting with Docker v.1.2.0, you can now edit /etc/hosts, /etc/hostname and /etc/resolve.conf in a running container. This is useful if you need to install bind or other services that might override one of those files.

Note, however, that changes to these files will not be saved by docker commit, nor will they be saved during docker run. That means they won't be saved in the image, nor will they persist when a container is restarted; they will only "stick" in a running container.