Access host (not vm) from inside container

I think you should be able to access the external IP of the Mac from a container with no issues, eg:

whale:docker justin$ ifconfig en0
en0: flags=8863<UP,BROADCAST,SMART,RUNNING,SIMPLEX,MULTICAST> mtu 1500
	ether ac:bc:32:cc:d0:89 
	inet6 fe80::aebc:32ff:fecc:d089%en0 prefixlen 64 scopeid 0x4 
	inet 172.16.10.16 netmask 0xffff0000 broadcast 172.16.255.255
	nd6 options=1<PERFORMNUD>
	media: autoselect
	status: active
whale:docker justin$ docker run alpine ping 172.16.10.16
PING 172.16.10.16 (172.16.10.16): 56 data bytes
64 bytes from 172.16.10.16: seq=0 ttl=37 time=0.221 ms
64 bytes from 172.16.10.16: seq=1 ttl=37 time=0.452 ms
^C
--- 172.16.10.16 ping statistics ---
2 packets transmitted, 2 packets received, 0% packet loss
round-trip min/avg/max = 0.221/0.336/0.452 ms

There is no standard name for this, and it will not work if you have no network access unfortunately, as your mac will not have an IP.

2 Likes

While that method works, it isn’t a perfect solution. IP addresses can change, and I don’t want my containers to stop working just because I’ve moved from the office to home.

VirtualBox provides the 10.0.2.2 address that maps from inside of a virtual machine back to the host. It would be nice if you could provide a similar mechanism.

Yes, that is correct, I can access via external IP when connected to a network. But as you said, this won’t work if unconnected.
And it also doesn’t help in the situation where, on the host, I want to run a service with a port bound to loopback only (such as python -m http.server --bind 127.0.0.1).

We are never going to be able to connect to the host loopback.

According to this stack overflow answer http://stackoverflow.com/questions/87442/virtual-network-interface-in-mac-os-x you can add aliases to loopback, which you might be able to use to give yourself a fixed address with eg ifconfig lo0 alias 172.16.123.1 - not tested this.

We currently cannot create an additional IP address due to some issues with OSX, and it is quite nice not having to interfere with the OSX network interfaces anyway, so we are less likely to break anything locally, or interfere with other software - even finding an IP address range that is unused can be difficult.

3 Likes

I understand not being able to connect to host loopback. I’m not even able to do that on Linux. It was basically an example of a use case-- I want to run a service on the host (python -m http.server, for example) where the ports are not exposed publicly, but yet I still want to be able to reach them (host + port) from within a container. Virtualbox + boot2docker/toolbox/etc. provide this mechanism with the 10.0.2.2 virtual interface. I can run a service on my OSX host without a publicly exposed port, and can reach it within any container without any special --add-host options. And I don’t have to be connected to a network at the time, and I can move my laptop from network to network without anything loss of functionality.

2 Likes

It appears to me like the whole point of this docker native is to make it feel like xyhve/hyper-v are not really there, and you’re clearly going through a lot of network and filesystem shenanigans to get us there, which is great.

But if I run a container and can’t simply “curl http://myhost.local” in some sane way, you’re cutting off how a whole segment of developers need to work.

Elaborate hacks and manual processes are essentially one-off and won’t be accepted as a mainstream solution, which makes it a barrier to entry as a every day workplace tool

It honestly can’t be that hard to accomplish this, we’re just talking about forwarding some packets at the end of the day… Compared to the filesystem work, this should be a piece of cake (I would imagine)

5 Likes

Using this methodology does seem to work, i.e.

sudo ifconfig lo0 alias 172.16.123.1
docker run -it --rm busybox ping 172.16.123.1

or with hostname mapping for app config convenience

docker run -it --rm --add-host=docker.local:172.16.123.1 busybox ping docker.local

But again… this is a manual process, making it a barrier to entry/adoption

If you’re concerned with coming up with non-conflicting IP schemes, then just make this an option in the settings panel. Give us a textbox to enter any IPv4 address into and automatically manage the lo0 alias on OS X.

This at least gives teams a consistent way to operate via a intuitive interface

4 Likes

@asherwin how would you achieve the same result on Docker running on a Linux machine?
Would it be assumed that the default gateway (from inside the container) is an address at which the host can be reached? E.g, within the container ip route | awk '/default/ { print $3 }' would equal the address on docker0.

At the risk of repeating myself, isn’t the point of docker native to make the end user feel like the Linux VM is “invisible”. The whole way this is setup is such that you never need to directly interact with that VM.

When running natively on Linux, the host your on is useful in other ways. You control this OS, configure it’s networking, maybe run non-containerized apps listening on various network ports that you might want your containers to talk to.

In this scenario, I’m saying that OS X/Windows should act in place of the actual Linux host in the “normal” setup… You’re abstracting away the xhyve/hyper-v VM entirely, so my default expectation would be that the containers would be able to communicate with the “host” , which for all intents and purposes here is OS X/Windows, not the underlying VM…

Given that would take a lot of network magic, I’d be happy if the above workaround with a lo0 alias was easily configured in a pretty settings panel, and the alias creation/removal was automatically managed by the app on start/shutdown, then that’s good enough to get packets to flow from the container → host in some sane, easily understood manner

6 Likes

Yes. It seems we’re in violent agreement here. We’ve gone a long way to try and achieve this and we want to make sure that the experience for Mac users is the same as the experience on Linux.

My question is perhaps more related to this statement:

You can’t curl myhost.local in Docker on Linux either, unless you manually inject something like with --add-host.

Imagine you are on a Linux desktop, running Docker and need to achieve the exact same use case as you’ve explained already. What do you do?

I’ve already outlined how I might do it… but I’d like to see how other’s might achieve the same thing. I’m hoping this will lead us in the direction of a solution that works for Docker for Mac users (i.e some networking magic we have to do on the VM to make this work).

I don’t think Docker for Mac managing loopback interfaces is a good idea for the reasons @justincormack mentioned + we’d be diverging from the Linux UX by adding a magic IP or Name that’s reachable from within all containers.

I can live with it as-is, but I guess I’m still stuck on the purpose of having docker feel native on OS X/Windows

I understand the programmer-centric thinking of all things same on all platforms, but let’s be realistic here, you’re doing docker native for OS X/Windows so developers can better seamlessly integrate docker into their everyday development workflows, which I applaud because I’ve been doing this manually VM’s since the early versions of docker with various hacks to get port’s and such exposed out to the host (OS X) etc.

I can’t imagine that even a stretch goal here is to ever expect people to run containers via docker native for real-world production usage. With that in mind, then what’s the harm in making something easier to use for developers. I’m not suggesting a docker cli UX change, we’re just talking about making some well-known IP (maybe off by default, and configurable in the docker native settings) that’s always-on that can be routed to from the docker containers, and this lives solely in the realm of the docker native VM and supporting application that manages it.

If your goal is to make the Linux VM invisible, then your “host running containers” is OS X, not Linux, and without the above there’s an obvious hurdle to making this a seamless integration into your development flow?

Here’s the scenario I envision:

Hey new guy, we all develop using docker native to support our apps, we run a custom nginx for authentication which must proxy all your apps, but we let you develop/run your apps from your native IDE. Go install docker native, and pull the latest custom nginx and set it up to proxy to the port you pick in your properties. Oh? What IP do you proxy to? Well that depends, every time you switch networks between office/home/vpn or get a new DHCP address you’re going to have to figure out your IP and go and re-configure your proxy, because, there’s no reasonable way for us to pre-configure things to “just work” for you. By the way you have to be online so your network interfaces are up.

Is this an elaborate example? Yes… but it’s one that’s already true for me, and I can easily envision lots of scenarios where you need containers to reach out to things running natively on your “host” (OS X, not Linux)

Again… I can live with this, I just think you’re missing an opportunity to make things easier on developers

2 Likes

The main issue is that we do not want to require root access on your Mac (we currently do use it, but only at initial install and we are trying to remove the use cases here too). Adding IP addresses requires root access, and as OSX does not have any kind of dummy interface it is kind of unclear what kind of interface to add. We can’t give it a name that containers would know about, so overall it does not seem terribly useful, and you probably may as well just use the lo0 hack.

The "lo0 hack" is basically the default behavior on Linux, as is root access. The default docker install creates the docker0 interface, and the container route/gateway to the host.
If this were done during install and documented, I think that would go a long way to making this feel closer to a standard Linux docker installation.
The documentation should note that this must be used in conjunction with pinata network=hostnet. With network=hostnet and the lo0 alias, I can achieve what I wanted, which is apps running on native OSX host that are reachable from containers, not reachable publicly, not subject to changes if I switch networks, and not subject to issues if I’m completely disconnected from any network.
With network=nat, I haven’t yet found a way to reach the host at all in any scenario, so I’m not sure what the point of it is-- unless the point is to have containers in a 100% isolated environment only reachable by each other.

2 Likes

I managed to get it working with network=nat. After I switched to nat, ifconfig showed a new network interface bridge100. I was able to successfully map hostnames to the bridge100 IP address and access services on the host.

@dteoh Care to share specifics?

$ IP=$(ifconfig | grep -A 3 bridge100 | grep inet | cut -d ’ ’ -f 2)
$ python3 -m http.server --bind $IP

In another terminal

$ IP=$(ifconfig | grep -A 3 bridge100 | grep inet | cut -d ’ ’ -f 2)
$ ping $IP
$ curl ${IP}:8000
$ docker run -it --rm busybox ping -c 1 $IP
$ docker run -it --rm --add-host=dh:${IP} buildpack-deps:curl curl --connect-timeout 5 dh:8000

The local commands work.
The docker commands fail.
The docker commands work if I use net=hostnet and the lo0 hack address.

I’d also be interested in this, as this is the only missing link preventing me from using docker for my dev environment.
Also, I’m a total newbie on Docker, thus it would be cool if the solution would be as less hackery as possible.

In my scenario, I want a container running php to be able to reach the database engine on the host (osx). Repeat for X projects. As what I want is the app services to be containers, but the DBs to be on my local for easier access, persistence, backup, handling, etc.

In one terminal:

$ docker version                                                                                                                                                                                                            08:19:09
Client:
 Version:      1.11.1
 API version:  1.23
 Go version:   go1.5.4
 Git commit:   5604cbe
 Built:        Wed Apr 27 00:34:20 2016
 OS/Arch:      darwin/amd64

Server:
 Version:      1.11.1
 API version:  1.23
 Go version:   go1.5.4
 Git commit:   8b63c77
 Built:        Tue May 10 10:39:20 2016
 OS/Arch:      linux/amd64
$ uname -a
Darwin laptop.local 14.5.0 Darwin Kernel Version 14.5.0: Thu Apr 21 20:40:54 PDT 2016; root:xnu-2782.50.3~1/RELEASE_X86_64 x86_64
$ python -m SimpleHTTPServer
Serving HTTP on 0.0.0.0 port 8000 ...

In another terminal:

$ ifconfig
bridge100: flags=8863<UP,BROADCAST,SMART,RUNNING,SIMPLEX,MULTICAST> mtu 1500
    options=3<RXCSUM,TXCSUM>
    ether 02:9a:9d:f9:9b:64
    inet 192.168.64.1 netmask 0xffffff00 broadcast 192.168.64.255
    Configuration:
        id 0:0:0:0:0:0 priority 0 hellotime 0 fwddelay 0
        maxage 0 holdcnt 0 proto stp maxaddr 100 timeout 1200
        root id 0:0:0:0:0:0 priority 0 ifcost 0 port 0
        ipfilter disabled flags 0x2
    member: en5 flags=3<LEARNING,DISCOVER>
            ifmaxaddr 0 port 11 priority 0 path cost 0
    nd6 options=1<PERFORMNUD>
    media: autoselect
    status: active
$ docker run --rm --add-host=dh:192.168.64.1 -it alpine ping dh                                                                                                                                                             08:14:14
PING dh (192.168.64.1): 56 data bytes
64 bytes from 192.168.64.1: seq=0 ttl=63 time=0.406 ms
64 bytes from 192.168.64.1: seq=1 ttl=63 time=0.473 ms
64 bytes from 192.168.64.1: seq=2 ttl=63 time=0.400 ms
64 bytes from 192.168.64.1: seq=3 ttl=63 time=0.427 ms
^C
--- dh ping statistics ---
4 packets transmitted, 4 packets received, 0% packet loss
round-trip min/avg/max = 0.400/0.426/0.473 ms
$  docker run --rm --add-host=dh:192.168.64.1 -it alpine wget http://dh:8000                                                                                                                                                 08:14:49
Connecting to dh:8000 (192.168.64.1:8000)
index.html           100% |*******************************|   178   0:00:00 ETA

I think the problem is that you are binding services to the bridge. My understanding is that the bridge is forwarding the packets onto localhost.

Anyway, I personally switched to the lo0 alias solution because:

  • I don’t have to modify the default docker settings
  • I am in control of the alias IP address which means not having to parse ifconfig output in scripts

I think aliasing loopback is the simplest solution. I also have the same requirements as you: accessing MySQL on the host. The only additional thing I had to do, which was MySQL specific, was not to use the root database user (because you have to explicitly grant extra permissions if connecting from outside of localhost) but to make a separate database user.

One problem with the lo0 hack is it’s not persistent-- you have to re-add the alias every time you reboot.
Also, the steps solution you provided @dteoh bind to all addresses, making the server public, which is one condition I’m trying to avoid, hence the reason for binding to the bridge address (or localhost, or alias, or whatever). If there’s another way to accomplish this, that’s what I’m looking for and open to.
When binding to a specific address, this (non-public, host reachable from containers) works on Linux and on Docker for Mac beta with lo0 hack address. On Linux, this is the only possible way (?) to accomplish this.

1 Like

I found this topic while I was searching for another issue which I discovered (I will mention later)

Host-networking works for me with host’s hostname and I’m switching daily in multiple networks (home, office, coffeeshops). The only case which I discovered was to set a dhcp-client-id because sometimes the dhcp-server overrides your host’s hostname.

sikei:docker christophkluge$ hostname
sikei.local

inside the docker container I can easily use

root@b6661fe0491f:/var/www# ping sikei
PING sikei (192.168.200.58): 56 data bytes
64 bytes from 192.168.200.58: icmp_seq=0 ttl=37 time=0.162 ms
64 bytes from 192.168.200.58: icmp_seq=1 ttl=37 time=0.242 ms


another issue which I see is that host latency is much slower then the latency to a different container. Since I’m not leaving the host mashine I would expect to have more close results in both scenarios

first example: roughly 3x times more latency

  1. ping host system
  2. ping another container

$ ping host -c 10
round-trip min/avg/max/stddev = 0.187/0.300/0.394/0.053 ms
$ ping mysql -c 10
round-trip min/avg/max/stddev = 0.076/0.092/0.120/0.000 ms

second example: roughly 10x slower (probably depends on the app + queries + amount of data)

  1. setup mysql database on your host system (import any kind of dump)
  2. setup mysql database inside a container (import the same dump)
  3. run your application pointing to host-system database
  4. run your application pointing to container-database

$ pinata diagnose
OS X: version 10.11.4 (build: 15E65)
Docker.app: version v1.11.1-beta13.1
Running diagnostic tests:
[OK] Moby booted
[OK] driver.amd64-linux
[OK] vmnetd
[OK] osxfs
[OK] db
[OK] slirp
[OK] menubar
[OK] environment
[OK] Docker
[OK] VT-x
Docker logs are being collected into /tmp/20160529-001229.tar.gz
Most specific failure is: No error was detected

2 Likes