The traffic hitting ping.archlinux.org has lately been exhausting its
default nf_conntrack_max limit of 64k. Bump it to 256k (which is also
the default limit found on systems with more than 4G of memory).

Suggested-by: Kristian Klausen <kristian@klausen.dk>

The default of 0.5 has proven insufficient on at least 3 boxes so far.

Fixes: 26f289b7 ("Capitalize the first letter of all task names")

ansible-lint 6.5.0 complains about:

  name: All names should start with an
        uppercase letter. (name[casing])

The glibc 2.35-6 package ships with the C.UTF-8 locale included which
means there is now a usable UTF-8 locale available by default.

en_US.UTF-8 will still be generated because PostgreSQL clusters are
created with that locale. Migrating the clusters to C.UTF-8 is
possible, but that requires dumping and recreating them.

These are used to signal the start of the document in a stream of many
documents. As Ansible only supports one YAML document per file this is
unnecessary. About a third of our YAML documents already lacked these.

Other than avoiding the conditional logic there's no reason to specify
/srv/ftp/pool/{packages,community} as additional cache directories for
all servers.

Restoring the default CacheDir value on machines without a local mirror
allows us to use pacdiff to perform a three-way merge of the few config
files which are easier to rebase this way.

Bash histories indicate this isn't being used anywhere other than
{build,gemini}.archlinux.org and gemini's filelist is so big that
locate becomes so slow that it's practically useless on this box.

There is a need for build servers to never build against outdated repo
databases, even with syncrepo providing a local mirror that is updated
every minute. To that effect, we adjust mirrorlist on build servers so
the first mirror is the tier0 mirror provided by gemini.

Keep the syncrepo role on build servers in order to have a local cache
of packages and avoid concurrent build jobs downloading the same files
causing them to be corrupted.

Finally, configure gemini to use its own repos (like other mirrors do).

Keeps us from having to deal with .pacnew files.

Run a self test every day at 02:00/03:00 and a long tests every
saturday.

sort + groupby fanciness by @klausenbusk.

Fix #392

Jan Alexander Steffens (heftig) authored 3 years ago and Kristian Klausen committed 3 years ago

- Allow killing of any cgroup if swap is about to run out. When swap is
  at least 90% full, oomd will kill a cgroup, tending to kill the cgroup
  with the highest swap usage.
- Allow killing cgroups of logged in users under severe memory pressure.
  When a cgroup is stalled waiting for memory at least 60% of the time,
  for at least 30 seconds, oomd will attempt to kill it or one of its
  descendant cgroups.

It has been applied upstream[1] and released with smartmontools 7.2[2].

[1] https://github.com/smartmontools/smartmontools/commit/61bc85ecadf9e5596b140380fed701a674e38a1b
[2] https://github.com/smartmontools/smartmontools/releases/tag/RELEASE_7_2

This reverts commit b4c97002.

Main motivation behind this is preparing for doing more automatic
upgrades. The need for ignoring kernel updates isn't clear either.

It was removed in:
d9fdafb0 ("Use archlinux-contrib over git submodule")
c380c429 ("Remove old task to symlink checkservices to /usr/local/bin")

Leonidas Spyropoulos authored 3 years ago and Kristian Klausen committed 3 years ago

This is shared between common and install_arch roles

Closed: #288

Signed-off-by: Leonidas Spyropoulos <artafinde@gmail.com>

Benedict Baskar authored 3 years ago and Kristian Klausen committed 3 years ago

Fix #376

Fixes: 678845af ("Add Kape server IPv6 addresses (fixes #230)")

It's been running out of swap during borg-backup and seems to get good
compression ratios; try upping the zram size to 100% of RAM (from 50%).

When both zswap and zram are active, zswap sits in front of zram and
treats it as a backing store. We just want to use zram and not zswap
disguising itself as such; disable the latter so we can enjoy useful
zramctl statistics.

Implemented as tmpfiles.d/zram.conf which disables zswap at runtime.

Restarting swap.target doesn't apply configuration changes; instead we
can restart systemd-zram-setup@zram0 which seems to do what we wanted.

Set "max-zram-size = none" to disable this unwanted limitation which
defaulted to creating zram-based swap with a maximum size of 4096MiB.

Fixes: dc8fa2bd ("common: Replace deprecated systemd-swap[1] with zram-generator")

Without hostname the role does not run.

Kristian Klausen authored 3 years ago and Jelle van der Waa committed 3 years ago

zram-generator defaults to 50% of the ram where systemd-swap defaults to
25% of the ram.

[1] https://lists.archlinux.org/pipermail/arch-dev-public/2021-May/030429.html

Leonidas Spyropoulos authored 3 years ago and Jelle van der Waa committed 3 years ago

Kernel's OOM killer is often too slow, as a result system ends up
thrashing heavily. Have a user-space oom killer helps.

Using default settings as described in
https://www.freedesktop.org/software/systemd/man/oomd.conf.html



Closes: #306

Signed-off-by: Leonidas Spyropoulos <artafinde@gmail.com>

A lot of our systems have a "hardware" watchdog (ex: all the Hetzner
cloud servers) so let's use it.

yaml: truthy value should be one of [false, true] (truthy)
yaml: wrong indentation: expected 4 but found 2 (indentation)
yaml: too few spaces before comment (comments)
yaml: missing starting space in comment (comments)
yaml: too many blank lines (1 > 0) (empty-lines)
yaml: too many spaces after colon (colons)
yaml: comment not indented like content (comments-indentation)
yaml: no new line character at the end of file (new-line-at-end-of-file)
load-failure: Failed to load or parse file
parser-error: couldn't resolve module/action 'hosts'. This often indicates a misspelling, missing collection, or incorrect module path.

It should also be installed before installing any packages with Ansible.

They were apparently controversial and more research is needed.

Do not use asterisk on network devices to prevent IP address collisions
on networks. Also use the right network mask for the assigned network.
For runner1 we need to ignore RAs since those routes don't work.