Repository navigation
Stop using podman for qm - #993
alexlarsson wants to merge 31 commits into
Conversation
|
The code looks good to me, even though this is a fairly large change. I couldn't find anything that caught my attention, particularly from a safety or potential side effects perspective. We will need time to investigate why the CI/CD pipeline is failing and to perform additional testing locally. These would be good candidates for follow-up Jira tickets. As a start, I would ask AI to fix the flake8 issues and shellcheck warnings here: Add your DCO too |
|
The bandit CI/CD failure seems false positive https://github.com/containers/qm/actions/runs/38773585735 From: |
|
I'll have a look at the failures later this week. But, the missing DCO is intentional, as I want to review things more before I sign off. |
|
I believe worth speak with @pypingou regarding if we want to mimic |
This is added next to the existing Makefile, but the intent is to replace it later, as we will eventually use it to build C code. Assisted-by: AI Signed-off-by: Alexander Larsson <alexl@redhat.com>
Use Meson for RPM builds and keep Makefile as a small developer wrapper. We move all the tests to CentOS 10, as that is the current target for QM. All backwards compat features for CentOS 9 are dropped. Assisted-by: AI Signed-off-by: Alexander Larsson <alexl@redhat.com>
This adds the new qmc binary, a C-based replacement for using podman to spawn the container. This binary replaces podman and netavark, and the primary use of it is to do the qm network setup and and to create the OCI container json config for crun. The actual running of the container is still done by crun, and the json file is similar. Assisted-by: AI Signed-off-by: Alexander Larsson <alexl@redhat.com>
Register the existing unittest and C fixtures with Meson and RPM checks, and build them in CentOS 10 CI. Assisted-by: AI Signed-off-by: Alexander Larsson <alexl@redhat.com>
Install qm.service and matching DNS/hosts watchers in place of the outer Quadlet. Also adds the needed sysctl config files. Assisted-by: AI Signed-off-by: Alexander Larsson <alexl@redhat.com>
Assisted-by: AI Signed-off-by: Alexander Larsson <alexl@redhat.com>
Move KVM/tun, sound, camera and TTY selections to qmc config drop-ins, instead of using quadlets with oci-hooks. This is fine now as the new config is much more expressible. Assisted-by: AI Signed-off-by: Alexander Larsson <alexl@redhat.com>
Pass qm-sysusers.conf explicitly to systemd-sysusers so enabling QM user namespaces does not process unrelated host account definitions. Assisted-by: AI
Set ping_group_range while preparing the host-owned network namespace, before crun enters the payload user namespace. Resolve payload root through the OCI GID mapping, retaining GID 0 without a user namespace. Remove the OCI sysctl that crun cannot set from the child namespace. Assisted-by: AI
Translate qmc-owned runtime files and host signal state into the payload namespace, including the private /run/secrets directory. Leave rootfs and writable etc/var mounts alone: their ownership is already shifted on disk. Check both cases in bundle preparation. Assisted-by: AI
Label qmc as a container runtime and relay non-terminal output through runtime-owned pipes that qm_t can inherit. Add -i/--interactive for stdin forwarding and -t/--tty for crun PTYs, with -t implying -i. Default stdin to /dev/null while preserving output and exit status. Use bounded nonblocking buffers so full output pipes or journal sockets cannot prevent signal forwarding. Cover stream integrity, file flags, backpressure, arguments and terminal selection in tests. Assisted-by: AI
Build the controller and QM nodes from CentOS Stream 10 and select the matching COPR repositories, as required by the native launcher. Assisted-by: AI
Move outer test nodes to 10.91.0.0/16 so their routes do not overlap QM at 10.90.0.0/24 or nested Podman at 10.89.0.0/16. Assisted-by: AI
Replace outer Podman exec, inspect and Quadlet configuration with qmc commands, native service state and JSON drop-ins across the existing integration suites. Preserve Podman for the nested workloads. Use the prepared writable trees for file copies, scope cleanup to test-owned drop-ins, and bound waits for nested service readiness. Assisted-by: AI
Add delegated fixtures for real crun startup, network/NAT behavior, cleanup and recovery, exec streams and signals, and pre-shifted filesystem ownership with mapped runtime mounts and ping sockets. Check systemd machine IDs, readiness and host signal path activation. Keep namespace and networking changes inside disposable test contexts; document the wrapper and retain explicit SELinux coverage limits. Assisted-by: AI
Add a bounded integration helper for a disposable machine with enforcing SELinux, prepared storage and a local nested image. Check service startup, exec, mounts, resolver updates, networking and nested Quadlets using one userns mode matching the fixture ownership. Assisted-by: AI
Replace the generated outer service directly, preserve activation during upgrades, and reject leftover Quadlet configuration at startup with migration instructions. Remove the host Podman requirement and obsolete CLI packages, enable native QM during setup, and preserve prepared filesystems on package removal. Assisted-by: AI
Preserve README topics while updating service, capability and resource examples. Add user-facing command, configuration and migration references, build qmc(1) and qmc.conf(5) from the same MkDocs sources, and port the existing usage, IPC, networking, subsystem and demo instructions. Assisted-by: AI Signed-off-by: Alexander Larsson <alexl@redhat.com>
This is basically a rewrite, so lets make that clear in the version. Signed-off-by: Alexander Larsson <alexl@redhat.com>
Write the controller socket address before restarting QM so a promptly starting agent reads the final configuration. Collect its journal through qmc exec on failure, alongside the host runtime journal. Signed-off-by: Alexander Larsson <alexl@redhat.com>
The range used to uncomment additionalimagestores did not stop at a closing bracket on its starting line. With the CentOS single-line default it also uncommented the next section description, leaving storage.conf invalid TOML and preventing nested Podman from starting. Stop on the current line for inline arrays and remove leading comment whitespace so the existing store insertion recognizes the setting. Signed-off-by: Alexander Larsson <alexl@redhat.com>
Remove the hook implementations, package, build jobs, and sanity checks. These are not needed anymore with the new podman-less qm. Signed-off-by: Alexander Larsson <alexl@redhat.com>
Drop the unused qm-is-ostree helper, the obsolete KVM tar exclusion list, and the empty host containers.conf.d installation. Remove parted from the runtime dependencies. The disk test helper installs it when partprobe is missing, without relying on the qm package to provide it. Signed-off-by: Alexander Larsson <alexl@redhat.com>
Define qm and qm-wayland as separate Packit packages, each with its own spec and source preparation action. Explicitly select the package on build, release, and test jobs, including the Wayland job inheriting the main Copr template. Keep all current test jobs attached to qm so a Wayland build cannot launch them with the wrong artifacts. Signed-off-by: Alexander Larsson <alexl@redhat.com>
Ignore the AIB test job until automotive-image-builder supports the qmc launcher. Signed-off-by: Alexander Larsson <alexl@redhat.com>
Non-terminal qmc exec keeps stderr separate from stdout. Capture both streams when checking setenforce and require a failing exit status as well as the expected permission denial, so command errors and actual success cannot be mistaken for a pass. Signed-off-by: Alexander Larsson <alexl@redhat.com>
The native launcher provides a PTY console for systemd to allow qm console output (from e.g. systemd) to go into the host journal. This means that the check that /dev/console doesn't exist breaks. So, accept a /dev/console, but only if it is a /dev/pts, and that its private (i.e. not the same as host /dev/console (in case that is a pts). Signed-off-by: Alexander Larsson <alexl@redhat.com>
Enable TMT's AVC check for the message-queue isolation test. Preserve the denial details when queue creation fails, rather than reporting only Permission denied. Signed-off-by: Alexander Larsson <alexl@redhat.com>
Give the libkrun Copr priority 90 so its libkrun and firmware are selected together instead of mixing newer AutoSD libkrun with older Copr firmware. Otherwise I'm, getting a conflict where it picks a newer libkrun from autosd than the one in the libkrun copr. Signed-off-by: Alexander Larsson <alexl@redhat.com>
This is currently broken with the latest crun due to: containers/crun#2487 Once that is fixed in cs10 we can re-enable this. Signed-off-by: Alexander Larsson <alexl@redhat.com>
There was a problem hiding this comment.
Sorry @alexlarsson, this pull request is too large to review
It changes 25,127 lines. Sourcery can review pull requests with up to 20,000 changed lines.
We have some intermittent failures in check_container_kvm_run_in_qm and check_qm_podman_quadlet_is_ok. This adds some more logging in case they fails so we can debug these better. Signed-off-by: Alexander Larsson <alexl@redhat.com>
This is a major change in how we start the qm container. Instead of using podman + quadlet, we're using a standard systemd qm.service file which starts "qmc launch". qmc is a small C program that generates the required OCI json file (similar to what podman would) in /run/qm/ and starts the container using crun. This lets qm start much faster, earlier and in ways more integrated with the system. The qmc app can be used in the installed image to e.g. run programs in the container (
qmc exec bash) and inspect the current status (qmc inspect). This app replaces the old qmctl, even though its not an exact 1-to-1 replacement.Interesting differences:
Note: while I tested and reviewed the main qmc work in another repo pretty carefully, the creation of this qm PR was mostly done by AI, so I marked this as draft, as I need to carefully review it. However, I want this in a PR early so people can look at it, and I can test out the CI changes.