With conditional restart (install-* tags), the homeserver entry for
Synapse fell through to `restart_necessary: true`, and the worker
entries carried no flag at all, which systemd_service_manager also
treats as "restart". Every install-all and install-service synapse
therefore restarted Synapse and all of its workers, even when nothing
had changed. Without workers, systemd also restarted most bridges and
bots, whose units require matrix-synapse.service.
Compute matrix_synapse_restart_necessary from the tasks that install
what Synapse runs: homeserver.yaml, the initial config and signing key,
the log config, labels and systemd service, the container image (pulled,
self-built or customized), each worker's config, labels and systemd
service, and the downloaded or cloned extension modules. Workers follow
the same decision, as they read the same files.
Synapse also mounts the appservice registration files written by other
roles. Add matrix_homeserver_app_service_registration_changed in
group_vars/matrix_servers, which checks the registered result of each
task installing one of the files listed in
matrix_homeserver_container_extra_arguments_auto, so that a changed
registration still restarts Synapse and its workers. The double-puppet
role did not register its result; it does now.
The Synapse image pull and three extension downloads all registered
their result as `result`. The downloads run first, so the pull's result
overwrote theirs before the restart decision could read them; they now
use role-prefixed names.
The ownership repair added in e5b8de8c2 is not sufficient on every Ansible privilege-escalation setup. Git may still see a different effective owner after the checkout has been recursively chowned, as demonstrated by #5065.
Pass an exact, task-scoped safe.directory setting to all 56 server-side git tasks. Keep the ownership repair as well, because it remains necessary for filesystem permissions. The two controller-side theme checkouts already use the same protection.
6b4b7647e fixed this for synapse-usage-exporter only. A checkout owned by
a different user makes the git task fail from then on, either with a
permission error or with git's dubious-ownership protection, until
someone removes the directory on the host by hand. It gets into that
state when the matrix user's uid changes (a server migration or a restore
onto a differently numbered user), when an earlier clone ran as another
user, or when someone runs git as root inside the checkout.
Every other role cloning a repository onto the server was open to the
same failure, so ensure the checkout's ownership recursively before
updating it at the remaining 54 sites.
The three matrix-synapse ext clones also gain force=yes. They were the
only on-server clones without it, which left a checkout that an
interrupted run had half-written wedged, instead of repaired on the next
run.
matrix-matrixto used to clone as root into a directory that nothing ever
chowned, unlike every other role. It now becomes the matrix user too.
The Element Web and SchildiChat Web theme checkouts live on the Ansible
controller, where correcting ownership is not ours to do, so they merely
mark the checkout as a safe directory for git.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
QEDeDandSlavi Pantaleev
