Docker and --userns-remap, how to manage volume pe

Question

In docker, files created inside containers tend to have unpredictable ownership while inspecting them from the host. The owner of the files on a volume is root (uid 0) by default, but as soon as non-root user accounts are involved in the container and writing to the file system, owners become more or less random from the host perspective.

It is a problem when you need to access volume data from the host using the same user account which is calling the docker commands.

Typical workarounds are

• forcing users uIDs at creation time in Dockerfiles (non portable)
• passing the UID of the host user to the docker run command as an environment variable and then running some chown commands on the volumes in an entrypoint script.

Both these solutions can give some control over the actual permissions outside the container.

I expected user namespaces to be the final solution to this problem. I have run some tests with the recently released version 1.10 and --userns-remap set to my desktop account. However, I am not sure that it can make file ownership on mounted volumes easier to deal with, I am afraid that it could actually be the opposite.

Suppose I start this basic container

docker run -ti -v /data debian:jessie /bin/bash
echo 'hello' > /data/test.txt
exit

And then inspect the content from the host :

ls -lh /var/lib/docker/100000.100000/volumes/<some-id>/_data/

-rw-r--r-- 1 100000 100000 6 Feb  8 19:43 test.txt

This number '100000' is a sub-UID of my host user, but since it does not correspond to my user's UID, I still can't edit test.txt without privileges. This sub-user does not seem to have any affinity with my actual regular user outside of docker. It's not mapped back.

The workarounds mentioned earlier in this post which consisted of aligning UIDs between the host and the container do not work anymore due to the UID->sub-UID mapping that occurs in the namespace.

Then, is there a way to run docker with user namespace enabled (for improved security), while still making it possible for the host user running docker to own the files generated on volumes?

Answer 1

If you can prearrange users and groups in advance, then it's possible to assign UIDs and GIDs in such specific way so that host users correspond to namespaced users inside containers.

Here's an example (Ubuntu 14.04, Docker 1.10):

1. Create some users with fixed numeric IDs:

   useradd -u 5000 ns1
   
   groupadd -g 500000 ns1-root
   groupadd -g 501000 ns1-user1
   
   useradd -u 500000 -g ns1-root ns1-root
   useradd -u 501000 -g ns1-user1 ns1-user1 -m
   

2. Manually edit auto-generated subordinate ID ranges in /etc/subuid and /etc/subgid files:

   ns1:500000:65536
   

   (note there are no records for ns1-root and ns1-user1 due to MAX_UID and MAX_GID limits in /etc/login.defs)

3. Enable user namespaces in /etc/default/docker:

   DOCKER_OPTS="--userns-remap=ns1"
   

   Restart daemon service docker restart, ensure /var/lib/docker/500000.500000 directory is created.

   Now, inside containers you have root and user1, and on the host -- ns1-root and ns1-user1, with matching IDs

   UPDATE: to guarantee that non-root users have fixed IDs in containers (e.g. user1 1000:1000), create them explicitly during image build.

Test-drive:

1. Prepare a volume directory

   mkdir /vol1
   chown ns1-root:ns1-root /vol1
   

2. Try it from a container

   docker run --rm -ti -v /vol1:/vol1 busybox sh
   echo "Hello from container" > /vol1/file
   exit
   

3. Try from the host

   passwd ns1-root
   login ns1-root
   cat /vol1/file
   echo "can write" >> /vol1/file
   

Not portable and looks like a hack, but works.

Answer 2

You can avoid permission problems by using the docker cp command.

Ownership is set to the user and primary group at the destination. For example, files copied to a container are created with UID:GID of the root user. Files copied to the local machine are created with the UID:GID of the user which invoked the docker cp command.

Here is your example switched to use docker cp:

$ docker run -ti -v /data debian:jessie /bin/bash
root@e33bb735a70f:/# echo 'hello' > /data/test.txt
root@e33bb735a70f:/# exit
exit
$ docker volume ls
DRIVER              VOLUME NAME
local               f073d0e001fb8a95ad8d919a5680e72b21a457f62a40d671b63c62ae0827bf93
$ sudo ls -l /var/lib/docker/100000.100000/volumes/f073d0e001fb8a95ad8d919a5680e72b21a457f62a40d671b63c62ae0827bf93/_data
total 4
-rw-r--r-- 1 100000 100000 6 Oct  6 10:34 test.txt
$ docker ps -a
CONTAINER ID        IMAGE               COMMAND             CREATED              STATUS                          PORTS               NAMES
e33bb735a70f        debian:jessie       "/bin/bash"         About a minute ago   Exited (0) About a minute ago                       determined_hypatia
$ docker cp determined_hypatia:/data/test.txt .
$ ls -l test.txt 
-rw-r--r-- 1 don don 6 Oct  6 10:34 test.txt
$ cat test.txt
hello
$ 

However, if you just want to read files out of a container, you don't need the named volume. This example uses a named container instead of a named volume:

$ docker run -ti --name sandbox1 debian:jessie /bin/bash
root@93d098233cf3:/# echo 'howdy' > /tmp/test.txt
root@93d098233cf3:/# exit
exit
$ docker cp sandbox1:/tmp/test.txt .
$ ls -l test.txt
-rw-r--r-- 1 don don 6 Oct  6 10:52 test.txt
$ cat test.txt
howdy
$ 

I find named volumes useful when I want to copy files into a container, as described in this question.