[arch-dev-public] [RFC] systemd service files policy
Hi guys,
We are hard at work adding systemd service files, and I thought it might be worth discussing a few guidelines for how to do it (should eventually end up in the wiki I suppose, but thought it would be best to start the discussion here).
I propose (based on input gathered on IRC) something like:
1) Use the upstream service files whenever they exist.
2) Don't do anything Arch-specific, if possible. This will maximize our chances of not having to change behavior in the future once the service file is provided by upstream. In particular avoid EnvironmentFile= especially if it points to the Arch-specific /etc/conf.d/*.
Not using an EnvironmentFile= is ok if:
2a) either the daemon has its own configuration file where the same settings can be specified. 2b) the default service file "just works" in the most common case. Users who want to change the behavior should then override the default service file.
If it is not possible to provide a sane default service file, I suppose we should discuss it on a case-by-case basis.
A few comments about service files (assuming we just want to preserve roughly the current behavior and not do anything fancy (if you want to get fancy, I'll assume you know what you are doing)):
3) If your service requires the network to be configured before it starts use After=network.target. Do not use Wants/Requires=network.target.
4) Use Type=forking, unless you know it is not necessary. Most daemons use double-forking to signal that they are ready, so to minimize the chance of problems it is probably safest to use this mode. To make sure that systemd is able to figure out which process is the main process, it is best to tell the daemon to write a pidfile and point systemd to it using PIDFile=. (If the daemon in question is dbus or socket activated, that's a different matter, but I guess that is only the case for the minority of daemons).
5) Our rc scripts do not support dependencies, but with systemd we should add them where necessary. The most typical case is that A requires the service B to be running before A is started. In that case add Requires=B and After=B to A. If the dependency is optional then add Wants=B and After=B instead. Dependencies are typically placed on services and not on targets (see 3).
Example of a simple conversion:
rc script =====
#!/bin/bash
. /etc/rc.conf . /etc/rc.d/functions
case "$1" in start) stat_busy "Starting NIS Server" /usr/sbin/ypserv if [ $? -gt 0 ]; then stat_fail else add_daemon ypserv stat_done fi ;; stop) stat_busy "Stopping NIS Server" killall -q /usr/sbin/ypserv if [ $? -gt 0 ]; then stat_fail else rm_daemon ypserv stat_done fi ;; restart) $0 stop sleep 1 $0 start ;; *) echo "usage: $0 {start|stop|restart}" esac
systemd service file ============
[Unit] Description=NIS/YP (Network Information Service) Server Requires=rpcbind.service After=network.target rpcbind.service
[Service] Type=forking PIDFile=/run/ypserv.pid ExecStart=/usr/sbin/ypserv
[Install] WantedBy=multi-user.target
On Thu, Aug 23, 2012 at 5:32 PM, Tom Gundersen teg@jklm.no wrote:
- Use the upstream service files whenever they exist.
This should be a SHOULD and not a MUST. Some service files can be poorly implemented or some deps are missing (e.g. your point 5) and be overloaded by maintainer.
Of course we want to push it upstream, like I've done with sslh, and now people add improvments upstream.
- Don't do anything Arch-specific, if possible. This will maximize
our chances of not having to change behavior in the future once the service file is provided by upstream. In particular avoid EnvironmentFile= especially if it points to the Arch-specific /etc/conf.d/*.
Not using an EnvironmentFile= is ok if:
I agree. I will gradually fix my packages.
- If your service requires the network to be configured before it
starts use After=network.target. Do not use Wants/Requires=network.target.
I agree.
- Use Type=forking, unless you know it is not necessary. Most daemons
use double-forking to signal that they are ready, so to minimize the chance of problems it is probably safest to use this mode. To make sure that systemd is able to figure out which process is the main process, it is best to tell the daemon to write a pidfile and point systemd to it using PIDFile=. (If the daemon in question is dbus or socket activated, that's a different matter, but I guess that is only the case for the minority of daemons).
I think we should _not_ use Type=Forking, except when it's the only choice. Systemd prefer not use forking type and it would be better to follow this.
Advantage to avoid forking can be read here [1] and in many others places. In short: - Disable forking may save each process to fork, setsid, wirte pidfiles, close stdin/stdout/stderr. - Save systemd logic to find main pid.
At first sight there is only pro but some peoples suggest to use forking for daemons which support foreground mode. The reason seems to be : systemd consider a unit started once launched when type=simple, whereas it consider a unit started when program exit when type=fork.
This can be useful *IF* daemon proceed in a specific order. In other words, it starts all it has to do and dameonize (double fork, etc) after. In most situations[2] daemons are not implemented like this and daemonize before starting all its services. So there is only cons to use Type=forking. An example is fcron[3] (We discuss of this with Dave on IRC last time).
So I think use Type=Forking is not the better recommendation. I would says : Try to use all other method before Type=Forking.
- Our rc scripts do not support dependencies, but with systemd we
should add them where necessary. The most typical case is that A requires the service B to be running before A is started. In that case add Requires=B and After=B to A. If the dependency is optional then add Wants=B and After=B instead. Dependencies are typically placed on services and not on targets (see 3).
I agree too.
[1] http://0pointer.de/blog/projects/systemd.html [2] It's a feeling not based on statistics. [3] https://bugs.archlinux.org/task/31248
-- Sébastien "Seblu" Luttringer www.seblu.net
On Thu, Aug 30, 2012 at 5:07 PM, Sébastien Luttringer seblu@seblu.net wrote:
On Thu, Aug 23, 2012 at 5:32 PM, Tom Gundersen teg@jklm.no wrote:
- Use the upstream service files whenever they exist.
This should be a SHOULD and not a MUST. Some service files can be poorly implemented or some deps are missing (e.g. your point 5) and be overloaded by maintainer.
Makes sense. Like with any other part of the packages, we might need to patch things until they are fixed upstream.
- Use Type=forking, unless you know it is not necessary. Most daemons
use double-forking to signal that they are ready, so to minimize the chance of problems it is probably safest to use this mode. To make sure that systemd is able to figure out which process is the main process, it is best to tell the daemon to write a pidfile and point systemd to it using PIDFile=. (If the daemon in question is dbus or socket activated, that's a different matter, but I guess that is only the case for the minority of daemons).
I think we should _not_ use Type=Forking, except when it's the only choice. Systemd prefer not use forking type and it would be better to follow this.
I agree that using anything else than Type=forking is preferable when possible, but correctness must come first. It is important that systemd is able to tell when a service is started, so that other services can order themselves After= it. This is achieved in the following ways:
Type=simple: not at all, it is assumed that this service may be considered "started" immediately. Either because it does not provide any communication channels for other services to connect to, or because those channels are set up using .socket's (the latter requires the daemon to contain specific code to deal with systemd, so is not so common yet).
Type=notify: the daemon will send a signal to systemd when it is ready, this requires systemd-specific code.
Type=dbus: the daemon is considered ready whenever a given dbus interface appears on the system bus.
Type=oneshot: this is considered started when it exits (so not applicable to long-running daemons).
Type=forking: this behaves exactly like traditional daemons: it is considered started when the first process has double-forked and exited, it might be necessary to use a PIDFile= so systemd knows which of the remaining processes is the main one.
As you can tell, all the types but "forking" depends on the daemons being implemented in a specific way, whereas Type=forking will work if and only if the old rc scripts worked correctly. So use something else if you know what you are doing, but as a general principle, I'd say that Type=forking is "safe" and anything else is an optimization we might leave for upstream to do, or get back to later once we have finished creating all the services.
In most situations[2] daemons are not implemented like this and daemonize before starting all its services. So there is only cons to use Type=forking.
If that is indeed the case, there is no point to using Type=forking. However, that would be a bug which also affects our old initscripts (and any other sysvinit-style init), so I would be surprised if it is very common.
-t
On 30 August 2012 23:40, Tom Gundersen teg@jklm.no wrote:
On Thu, Aug 30, 2012 at 5:07 PM, Sébastien Luttringer seblu@seblu.net wrote:
On Thu, Aug 23, 2012 at 5:32 PM, Tom Gundersen teg@jklm.no wrote:
- Use the upstream service files whenever they exist.
This should be a SHOULD and not a MUST. Some service files can be poorly implemented or some deps are missing (e.g. your point 5) and be overloaded by maintainer.
Makes sense. Like with any other part of the packages, we might need to patch things until they are fixed upstream.
What concerns me is the "overloaded by maintainer" part. For e.g., what if the maintainer disagrees with one single approach that appears to come from our distribution? Do we then have to keep our own service file "fork"?
-- GPG/PGP ID: C0711BF1
On Thu, Aug 30, 2012 at 5:52 PM, Rashif Ray Rahman schiv@archlinux.org wrote:
What concerns me is the "overloaded by maintainer" part. For e.g., what if the maintainer disagrees with one single approach that appears to come from our distribution? Do we then have to keep our own service file "fork"?
I'd say we should treat it as we always do: follow upstream whenever possible and only patch critical bugs.
-t
On Thu, Aug 30, 2012 at 5:40 PM, Tom Gundersen teg@jklm.no wrote:
On Thu, Aug 30, 2012 at 5:07 PM, Sébastien Luttringer seblu@seblu.net wrote:
On Thu, Aug 23, 2012 at 5:32 PM, Tom Gundersen teg@jklm.no wrote: I think we should _not_ use Type=Forking, except when it's the only choice. Systemd prefer not use forking type and it would be better to follow this.
I agree that using anything else than Type=forking is preferable when possible, but correctness must come first.
That was I think we should underline in our wiki.
It is important that systemd is able to tell when a service is started, so that other services can order themselves After= it.
Sure.
participants (3)
-
Rashif Ray Rahman
-
Sébastien Luttringer
-
Tom Gundersen