Skip to content

Docker: allow to be used from NIPA with rootless podman - #139

Merged
matttbe merged 8 commits into
linux-netdev:mainfrom
matttbe:docker-sft-nipa
Nov 17, 2026
Merged

matttbe merged 8 commits into
linux-netdev:mainfrom
matttbe:docker-sft-nipa

Conversation

@matttbe

@matttbe matttbe commented Nov 17, 2026

Copy link
Copy Markdown
Member

A few modifications to run the docker images from nipa with Podman

This message is useful when the script is used interactively, but it is
just noise when the output is parsed, e.g. by contest/remote/lib/vm.py
when 'vng' is executed from the container.

Signed-off-by: Matthieu Baerts <matttbe@kernel.org>
Directories passed as arguments are mounted, but not when they are
passed as values, e.g. 'make O=<dir>', 'INSTALL_PATH=<dir>', or
'--opt=<dir>'.

That's needed to wrap 'make' with this script when executing
contest/remote/vmksft-p.py: get_prog_list() installs the selftests in a
temp dir created on the host with 'INSTALL_PATH=<dir>', then reads the
generated kselftest-list.txt file from there. Without this mount, the
file would only be created inside the container.

Signed-off-by: Matthieu Baerts <matttbe@kernel.org>
To easily execute some commands from the selftests container instead of
the host, e.g. 'make' and 'vng' when executing contest/remote/vmksft-p.py
to use the toolchain, QEMU and virtme-ng from the container: symlinks
named after these commands and pointing to this wrapper can be placed in
a directory at the beginning of $PATH.

This wrapper is not a symlink to a common script, because it would then
be resolved by 'realpath', and the wrong image would be used.

Signed-off-by: Matthieu Baerts <matttbe@kernel.org>
When the Docker client is killed with SIGKILL, the container is not
stopped. This is what contest/remote/lib/vm.py does when a VM doesn't
stop: all its processes are killed with SIGKILL. When 'vng' is executed
from a container, QEMU then continues to run in the background, using
CPU and memory resources.

Such signal cannot be trapped, nor forwarded by the client. Instead, a
detached process -- so not part of the killed process tree -- waits for
the client to be gone, then removes the container, identified by a new
unique name. This process also removes the container if it is asked to
stop, e.g. when all processes of a systemd service are terminated.

This watchdog is not started if one of the commands it uses is wrapped
with wrapper.sh: it would call itself recursively. A warning is printed
instead.

While at it, the client is now started with 'exec': the signals sent to
the script's PID are then received by the client, which forwards them
to the container. '--init' is now used to avoid having these signals
ignored by the command, executed as PID 1 in the container otherwise.

Signed-off-by: Matthieu Baerts <matttbe@kernel.org>
The extra paths from '[env] paths' were appended to $PATH: they could
then not be used to override commands from the host, e.g. to execute
'make' and 'vng' from a container via docker/selftests/wrapper.sh.

Prepend them instead, like it is already done in contest/remote/exec.py.

Signed-off-by: Matthieu Baerts <matttbe@kernel.org>
The output of the container is sent to the attached client, but it is
also sent to a log driver by default. Depending on the config, it is
then written a second time to the disk -- 'json-file', without size
limit by default -- or to the journal -- 'journald', the default with
Podman -- where it is kept after the removal of the container.

That's not needed here: the output is only used via the client, and the
container is removed at the end. That's also not wanted when executing
'vng' from the container: the whole serial console of each VM would be
duplicated.

Signed-off-by: Matthieu Baerts <matttbe@kernel.org>
With rootless Podman, the user's UID is mapped to root in the container
by default, and '-u $(id -u)' is then mapped to another UID on the host,
which cannot write to the mounted directories, e.g. the kernel tree.

'--userns=keep-id' fixes that by keeping the same UID/GID in the
container. Podman uses the PODMAN_USERNS environment variable as the
default value for this option, while Docker ignores it: the same script
can then be used with both, e.g. with Podman via the podman-docker
package providing the 'docker' command.

This is only done for non-root users, not to change the behaviour with
rootful Podman.

Signed-off-by: Matthieu Baerts <matttbe@kernel.org>
The scripts can be used with Podman, via the 'docker' command provided
by the podman-docker package. With rootless Podman, '--userns=keep-id'
is set by the scripts, but it has to be added when using Podman
directly, to be able to write to the mounted directories.

When used from a systemd service with 'User=', lingering has to be
enabled for this user, and the unit has to set its runtime directory:
Podman can then use the user's systemd instance, e.g. to manage the
containers' cgroups, without printing warnings at each call.

Signed-off-by: Matthieu Baerts <matttbe@kernel.org>
@matttbe

matttbe commented Nov 17, 2026

Copy link
Copy Markdown
Member Author

Exceptionally merging without review to deploy this workaround.

@matttbe
matttbe merged commit 28b814d into linux-netdev:main Nov 17, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant