Linux – LXC: Any security difference between root and end-user owned unprivileged containers

Since the release of 0.9 Docker has dropped LXC and uses its own execution environment, libcontainer. Your question's a bit old but I guess my answer still applies the version you are using.

Quick Answer: To understand the permissions of volumes, you can take the analogy of mount --bind Host-Dir Container-Dir. So to fulfill your requirement you can use any traditional methods for managing permissions. I guess ACL is what you need.

Long Answer: So as in your example we have a container named dock with a volume /data.

docker run -tid --name dock -v /usr/container/Databases/:/data \
    centos:latest /bin/bash

Inside the container our MySQL server has been configured to use the /data as its data directory. So we have our databases in the /data inside the container. And outside the container on the Host OS, we have mounted this /data volume from /usr/container/Databases/ and we assign a normal user bob to take backups of the databases. From the host machine we'll configure ACLs for user bob.

useradd -u 3000 bob
usermod -R o=--- /usr/container/Databases/
setfacl -R -m u:bob:rwx /usr/container/Databases/
setfacl -R -d -m u:bob:rwx /usr/container/Databases/

To test it out lets take a backup with user bob.

su - bob
tar -cvf container-data.tar /usr/container/Databases/

And tar will list out and you can see that our user was able to access all the files.

Now from inside the container if you check with getfacl you will notice that instead of bob it shows 3000. This is because the UID of bob is 3000 and there is no such user in the container so it simply displays the UID it receives from the meta data. Now if you create a user in your container with useradd -u 3000 bob you will notice that now the getfacl shows the name bob instead of 3000.

Summary: So the user permissions you assign from either inside or outside the container reflects to both environments. So to manage the permissions of volumes, UIDs in host machine must be different from the UIDs in the container.

I know now how to do this. If you can't follow this explanation, please ask back, but also make sure you have read up on userns in the readings I am giving at the bottom

Preliminary assumptions

I'll stick with the following assumptions, extended from what I have from your question:

1. host has a user1 and a user2, if an information isn't specific to one, we'll use userX
2. the container will be named by a variable which we will render as $container
3. home folders for user1 and user2 will be given in the notation known from Bash as ~user1 and ~user2.
4. we'll assume the subordinate UID and GID ranges to be 100000..165536 for user1 and 200000..265536 for user2 just for brevity
5. the root FS folder for $container will be rendered as $rootfs, regardless of where it will end up (~userX/.local/share/lxc/$container/rootfs)
6. container configuration is by default in ~userX/.local/share/lxc/$container/config

Moving the container

There are two relevant pieces of data that govern the userns containers:

1. owner and group for the files/folders of the folders comprising the $container
2. the subordinate UIDs and GIDs assigned in two places: /etc/sub{uid,gid} for the user account (manipulated via usermod --{add,del}-sub-{uid,gid}s) and lxc.id_map in the $container configuration (~userX/.local/share/lxc/$container/config) respectively
   • I don't know for certain whether it is possible to define different ranges in the container configuration for each container. E.g. if the host user userX has 65536 subordinate GIDs and UIDs, it might be possible to assign 5000 to 65 different containers, but I haven't tested that hypothesis.
   • it is certain, though, that this setting communicates to LXC which are the valid ranges for GID and UID in the child namespace.

So the gist is really that you need to make sure that the file/folder owner and group for the container match the configuration, which in turn has to be a valid subset of the host subordinate GIDs/UIDs assigned to user1 and user2 respectively.

If you're using Bash, for example, you can use $((expression)) for arithmetic expressions and let to assign arithmetic expressions to variables. This is mighty useful if you know a base value (100000 and 200000 respectively) and the GID/UID for the "inside" users.

The main points are:

1. it's possible
2. either the capability CAP_CHOWN or superuser rights are required

Here's a script that will probably need some more honing (example: migration from root-created container to unprivileged), but it works for me for the purpose:

#!/usr/bin/env bash

function syntax
{
    echo "SYNTAX: ${0##*/} <from-user> <to-user> <container-name>"
    [[ -n "$1" ]] && echo -e "\nERROR: ${1}."
    exit 1
}

# Checks
[[ -n "$1" ]] || syntax "<from-user> is not set"
[[ -n "$2" ]] || syntax "<to-user> is not set"
[[ -n "$3" ]] || syntax "<container-name> is not set"
[[ "$UID" -eq "0" ]] || syntax "${0##*/}" "You must be superuser to make use of this script"
# Constants with stuff we need
readonly USERFROM=$1
readonly USERTO=$2
shift; shift
readonly CONTAINER=${1:-*}
LXCLOCAL=".local/share/lxc"
readonly HOMEFROM=$(eval echo ~$USERFROM)
readonly HOMETO=$(eval echo ~$USERTO)
readonly LXCFROM="$HOMEFROM/$LXCLOCAL"
readonly LXCTO="$HOMETO/$LXCLOCAL"
readonly GIDBASEFROM=$(awk -F : "\$1 ~/$USERFROM/ {print \$2}" /etc/subgid)
readonly UIDBASEFROM=$(awk -F : "\$1 ~/$USERFROM/ {print \$2}" /etc/subuid)
readonly GIDSIZEFROM=$(awk -F : "\$1 ~/$USERFROM/ {print \$3}" /etc/subgid)
readonly UIDSIZEFROM=$(awk -F : "\$1 ~/$USERFROM/ {print \$3}" /etc/subuid)
readonly GIDBASETO=$(awk -F : "\$1 ~/$USERTO/ {print \$2}" /etc/subgid)
readonly UIDBASETO=$(awk -F : "\$1 ~/$USERTO/ {print \$2}" /etc/subuid)
readonly GIDSIZETO=$(awk -F : "\$1 ~/$USERTO/ {print \$3}" /etc/subgid)
readonly UIDSIZETO=$(awk -F : "\$1 ~/$USERTO/ {print \$3}" /etc/subuid)
unset LXCLOCAL
# More checks
[[ -d "$LXCFROM" ]] || syntax "Could not locate '$LXCFROM'. It is not a directory as expected"
[[ -e "$LXCTO" ]] && syntax "Destination '$LXCTO' already exists. However, it must not"
for i in GIDBASEFROM UIDBASEFROM GIDBASETO UIDBASETO; do
    (($i > 0)) || syntax "Could not determine base/offset of subordinate UID/GID range"
done
for i in GIDSIZEFROM UIDSIZEFROM GIDSIZETO UIDSIZETO; do
    (($i > 0)) || syntax "Could not determine length of subordinate UID/GID range"
done

echo "Going to migrate container: $CONTAINER"
echo -e "\tfrom user $USERFROM ($HOMEFROM): subUID=${UIDBASEFROM}..$((UIDBASEFROM+UIDSIZEFROM)); subGID=${GIDBASEFROM}..$((GIDBASEFROM+GIDSIZEFROM))"
echo -e "\tto user $USERTO ($HOMETO): subUID=${UIDBASETO}..$((UIDBASETO+UIDSIZETO)); subGID=${GIDBASETO}..$((GIDBASETO+GIDSIZETO))"
while read -p "Do you want to continue? (y/N) "; do
    case ${REPLY:0:1} in
        y|Y)
            break;
            ;;
        *)
            echo "User asked to abort."
            exit 1
            ;;
    esac
done

# Find the UIDs and GIDs in use in the container
readonly SUBGIDSFROM=$(find -H "$LXCFROM" -printf '%G\n'|sort -u)
readonly SUBUIDSFROM=$(find -H "$LXCFROM" -printf '%U\n'|sort -u)

# Change group
for gid in $SUBGIDSFROM; do
    let GIDTO=$(id -g "$USERTO")
    if ((gid == $(id -g "$USERFROM"))); then
        echo "Changing group from $USERFROM ($gid) to $USERTO ($GIDTO)"
        find -H "$LXCFROM/$CONTAINER" -gid $gid -exec chgrp $GIDTO {} +
    elif ((gid >= GIDBASEFROM )) && ((gid <= GIDBASEFROM+GIDSIZEFROM)); then
        let GIDTO=$((gid-GIDBASEFROM+GIDBASETO))
        echo "Changing group $gid -> $GIDTO"
        find -H "$LXCFROM/$CONTAINER" -gid $gid -exec chgrp $GIDTO {} +
    else
        echo "ERROR: Some file/folder inside '$LXCFROM/$CONTAINER' has a group not assigned to $USERFROM (assigned subordinate GIDs)."
        echo -e "Use:\n\tfind -H '$LXCFROM/$CONTAINER' -gid $gid\nto list those files/folders."
        exit 1
    fi
done

# Change owner
for uid in $SUBUIDSFROM; do
    let UIDTO=$(id -u "$USERTO")
    if ((uid == $(id -u "$USERFROM"))); then
        echo "Changing owner from $USERFROM ($uid) to $USERTO ($UIDTO)"
        find -H "$LXCFROM/$CONTAINER" -uid $uid -exec chown $UIDTO {} +
    elif ((uid >= UIDBASEFROM )) && ((uid <= UIDBASEFROM+UIDSIZEFROM)); then
        let UIDTO=$((uid-UIDBASEFROM+UIDBASETO))
        echo "Changing owner $uid -> $UIDTO"
        find -H "$LXCFROM/$CONTAINER" -uid $uid -exec chown $UIDTO {} +
    else
        echo "ERROR: Some file/folder inside '$LXCFROM/$CONTAINER' has an owner not assigned to $USERFROM (assigned subordinate UIDs)."
        echo -e "Use:\n\tfind -H '$LXCFROM/$CONTAINER' -uid $uid\nto list those files/folders."
        exit 1
    fi
done
mv "$LXCFROM/$CONTAINER" "$LXCTO/" || { echo "ERROR: failed to move to destination: ${LXCTO}/${CONTAINER}."; exit 1; }

In addition to the license terms of the StackExchange network, I am putting this into the public domain. So reuse and modify for whatever purpose, but it comes without any warranty and I must not be held liable for its use or abuse.

SYNTAX: lxc-reassign-userns.sh <from-user> <to-user> <container-name>

It assumes find, sort, uniq, awk (mawk and gawk should work), id, bash, chown, chmod and so on to be available and to understand all the command line switches it is using. For Bash readonly and let and arithmetic expressions are assumed to be understood. For find is assumes + is a valid terminator for the -exec action.

This list is probably not complete.

Backups

Yes, you can make backups and restore them elsewhere, as long as you also adjust the file owner and group accordingly.

However, assuming you use something like tar, there's a caveat: tar will ignore sockets, so $rootfs/dev/log will pose an issue - others may also create a similar issue.

Resources:

• subuid(5)
• subgid(5)
• newuidmap(1)
• newgidmap(1)
• user_namespaces(7)