Users and groups

This is one of the more complex topics. Generally speaking, bootc has nothing to do directly with configuring users or groups; it is a generic OS update/configuration mechanism. (There is currently just one small exception in that bootc install has a special case --root-ssh-authorized-keys argument, but it's very much optional).

Generic base images

Commonly OS/distribution base images will be generic, i.e. without any configuration. It is very strongly recommended to avoid hardcoded passwords and ssh keys with publicly-available private keys (as Vagrant does) in generic images.

Injecting SSH keys via systemd credentials

The systemd project has documentation for credentials which can be used in some environments to inject a root password or SSH authorized_keys. For many cases, this is a best practice.

At the time of this writing this relies on SMBIOS which is mainly configurable in local virtualization environments. (qemu).

Injecting users and SSH keys via cloud-init, etc.

Many IaaS and virtualization systems are oriented towards a "metadata server" (see e.g. AWS instance metadata) that are commonly processed by software such as cloud-init or Ignition or equivalent.

The base image you're using may include such software, or you can install it in your own derived images.

In this model, SSH configuration is managed outside of the bootable image. See e.g. GCP oslogin for an example of this where operating system identities are linked to the underlying Google accounts.

Adding users and credentials via custom logic (container or unit)

Of course, systems like cloud-init are not privileged; you can inject any logic you want to manage credentials via e.g. a systemd unit (which may launch a container image) that manages things however you prefer. Commonly, this would be a custom network-hosted source. For example, FreeIPA.

Another example in a Kubernetes-oriented infrastructure would be a container image that fetches desired authentication credentials from a CRD hosted in the API server. (To do things like this it's suggested to reuse the kubelet credentials)

System users and groups (added via packages, etc)

It is common for packages (deb/rpm/etc) to allocate system users or groups as part of e.g apt|dnf install <server package> such as Apache or MySQL, and this is often done by directly invoking useradd or groupadd as part of package pre/post installation scripts.

With theshadow-utils implementation of useradd and the default glibc files this will result in changes to the traditional /etc/passwd and /etc/shadow files as part of the container build.

System drift from local /etc/passwd modifications

When the system is initially installed, the /etc/passwd in the container image will be applied and contain desired users.

By default (without etc = transient, see below), the /etc directory is machine-local persistent state. If subsequently /etc/passwd is modified local to the machine (as is common for e.g. setting a root password) then any new changes in the container image (such as users from new packages) will not appear on subsequent updates by default (they will be in /usr/etc/passwd instead - the default image version).

The general best fix for this is to use systemd-sysusers instead of allocating a user/group at build time at all.

Using systemd-sysusers

See systemd-sysusers. For example in your derived build:

COPY mycustom-user.conf /usr/lib/sysusers.d

A key aspect of how this works is that sysusers will make changes to the traditional /etc/passwd file as necessary on boot instead of at build time. If /etc is persistent, this can avoid uid/gid drift (but in the general case it does mean that uid/gid allocation can depend on how a specific machine was upgraded over time).

User and group home directories and /var

For systems configured with persistent /home → /var/home, any changes to /var made in the container image after initial installation will not be applied on subsequent updates. If for example you inject /var/home/someuser/.ssh/authorized_keys into a container build, existing systems will not get the updated authorized keys file.

Using DynamicUser=yes for systemd units

For "system" users it's strongly recommended to use systemd DynamicUser=yes where possible.

This is significantly better than the pattern of allocating users/groups at "package install time" (e.g. Fedora package user/group guidelines) because it avoids potential UID/GID drift (see below).

Using systemd JSON user records

See JSON user records. Unlike sysusers, the canonical state for these live in /usr - if a subsequent image drops a user record, then it will also vanish from the system - unlike sysusers.d.

nss-altfiles

The nss-altfiles project (long) predates systemd JSON user records. It aims to help split "system" users into /usr/lib/passwd and /usr/lib/group. It's very important to understand that this aligns with the way the OSTree project handles the "3 way merge" for /etc as it relates to /etc/passwd. Currently, if the /etc/passwd file is modified in any way on the local system, then subsequent changes to /etc/passwd in the container image will not be applied.

Some base images may have nss-altfiles enabled by default; this is currently the case for base images built by rpm-ostree.

Commonly, base images will have some "system" users pre-allocated and managed via this file again to avoid uid/gid drift.

In a derived container build, you can also append users to /usr/lib/passwd for example. (At the time of this writing there is no command line to do so though).

Typically it is more preferable to use sysusers.d or DynamicUser=yes.

Machine-local state for users

At this point, it is important to understand the filesystem layout - the default is up to the base image.

The default Linux concept of a user has data stored in both /etc (/etc/passwd, /etc/shadow and groups) and /home. The choice for how these work is up to the base image, but a common default for generic base images is to have both be machine-local persistent state. In this model /home would be a symlink to /var/home/someuser.

Injecting users and SSH keys via at system provisioning time

For base images where /etc and /var are configured to persist by default, it will then be generally supported to inject users via "installers" such as Anaconda (interactively or via kickstart) or any others.

Typically generic installers such as this are designed for "one time bootstrap" and again then the configuration becomes mutable machine-local state that can be changed "day 2" via some other mechanism.

The simple case is a user with a password - typically the installer helps set the initial password, but to change it there is a different in-system tool (such as passwd or a GUI as part of Cockpit, GNOME/KDE/etc).

It is intended that these flows work equivalently in a bootc-compatible system, to support users directly installing "generic" base images, without requiring changes to the tools above.

Transient home directories

Many operating system deployments will want to minimize persistent, mutable and executable state - and user home directories are that

But it is also valid to default to having e.g. /home be a tmpfs to ensure user data is cleaned up across reboots (and this pairs particularly well with a transient /etc as well):

In order to set up the user's home directory to e.g. inject SSH authorized_keys or other files, a good approach is to use systemd tmpfiles.d snippets:

f~ /home/someuser/.ssh/authorized_keys 600 someuser someuser - <base64 encoded data>

which can be embedded in the image as /usr/lib/tmpfiles.d/someuser-keys.conf.

Or a service embedded in the image can fetch keys from the network and write them; this is the pattern used by cloud-init and afterburn.

UID/GID drift

Any invocation of useradd or groupadd that does not allocate a fixed UID/GID may be subject to "drift" in subsequent rebuilds by default.

One possibility is to explicitly force these user/group allocations into a static state, via systemd-sysusers (per above) or explicitly adding the users with static IDs before a dpkg/RPM installation script operates on it:

RUN <<EORUN
set -xeuo pipefail
groupadd -g 10044 mycustom-group
useradd -u 10044 -g 10044 -d /dev/null -M mycustom-user
dnf install -y mycustom-package.rpm
bootc container lint
EORUN

Ultimately the /etc/passwd and similar files are a mapping between names and numeric identifiers. A problem then becomes when this mapping is dynamic and mixed with "stateless" container image builds.

For example today the CentOS Stream 9 postgresql package allocates a static uid of 26.

This means that

RUN dnf -y install postgresql

will always result in a change to /etc/passwd that allocates uid 26 and data in /var/lib/postgres will always be owned by that UID.

However in contrast, the cockpit project allocates a floating cockpit-ws user.

This means that each container image build (without additional work, unlike the example at the begining of this section),may (due to RPM installation ordering or other reasons) result in the uid changing.

This can be a problem if that user maintains persistent state. Such cases are best handled by being converted to use sysusers.d (see Fedora change) - or again even better, using DynamicUser=yes (see above).

tmpfiles.d use for setting ownership

Systemd's tmpfiles.d provides a way to define files and directories in a way that will be processes at startup as needed. One way to work around SELinux security context and user or group ownership of a directory or file can be by using the z or Z directives.

These directives will adjust the access mode, user and group ownership and the SELinux security context as stated on the doc linked above.

For example, if we need /var/lib/my_file.conf to be part of the tss group but owned by root we could create a tmpfiles.d entry with:

+z /var/lib/my_file 0640 root tss -

The Linux Foundation® (TLF) has registered trademarks and uses trademarks. For a list of TLF trademarks, see Trademark Usage.