Thank you very much for this, it greatly helped me understand how this all works.

I have therefore solved this issue with the following simple solution:

[root@example ~]# cat /etc/systemd/system/teamd\@.service.d/before_network.conf
[Unit]
Before=network.target

I will post this in RH bugzilla as a potential line to be included in the distro teamd unit file.

Ben


On 02/26/2015 12:52 PM, Andrei Borzenkov wrote:
В Thu, 26 Feb 2015 11:56:45 -0500
Benjamin Rose <benr...@math.princeton.edu> пишет:

So given the distro unit file for teamd on RHEL7:

[Unit]
Description=Team Daemon for device %I
[Service]
BusName=org.libteam.teamd.%i
ExecStart=/usr/bin/teamd -U -D -t %i -f /run/teamd/%i.conf

Should I just make something like this...?

/etc/systemd/system/teamd@.service.d/require_network.conf:
Before=network.service

Perhaps I am misunderstanding the relationship between network.service
and network.target?

network.target is common synchronization point. Services that need to
be started after network can order itself after it. Services that
implement network should order itself before it. This frees individual
consumers of dependencies on actual implementation.

                     Really, network.service is what starts the teamd
processes.
This is network.service in your case. In other cases it could be
NetworkManager.service, or systemd-networkd.service or wicked.service.

           The question is how to make sure systemd knows to stop
pacemaker before stopping all of the teamd processes.
By ordering pacemaker after teamd services. But if you want to be
independent of actual implementation, pacemaker should come After
network.target and teamd (or whatever) should come Before
network.target. If teamd itself is instantiated template, just include
Before in template and every instance will automatically inherit it.

                                                      I would have
thought since network.service generated the teamd instances, it would
have known to keep them alive until all services requiring network were
finished.

I'm not sure what "it" in the above sentence refers to. If you mean
"systemd" then it only respects After and Before relations. This is
even documented :)

Thanks,
Ben

On 02/25/2015 10:38 PM, Andrei Borzenkov wrote:
В Wed, 25 Feb 2015 18:42:47 -0500
Benjamin Rose <benr...@math.princeton.edu> пишет:

Hello all,

I hope this is the right place for this inquiry. I was noticing
extremely slow reboot times on three of my hosts running
corosync/pacemaker for shared storage. I enabled the systemd debug logs,
and found that pacemaker was attempting to communicate with it's peers
to notify of the shutdown, and failed to do so as networking was not
functioning. This was odd because networking was fine during normal
operation and the distribution's unit file has the proper "Requires" for
networking. Eventually it hit the distro "TimeoutStopSec=30m" and forced
the reboot, hence the slow reboot times.

In the debug logs, I found that it was all because I am using teamd
(partially related - through ifcfg files using network.service), and
teamd was being killed long before pacemaker got the chance to send its
closing messages. So, I fixed the problem in my implementation with this
little bit:

[root@myhost ~]# cat
/etc/systemd/system/pacemaker.service.d/require_teamd.conf
[Unit]
After=teamd@team_pub.service
After=teamd@team_priv.service
Requires=teamd@team_pub.service
Requires=teamd@team_priv.service

But, soon I will be changing a lot about my networking. I'll be using
puppet to deploy a few more teams and bridges on this host. So, my
question comes down to - is there a way to accomplish something like this:

Requires=teamd@*.service
After=teamd@*.service

to include all running instances? I know this makes no sense in an
xinetd-type situation on bootup, when instances will be created
on-demand, but it does make perfect sense on a shutdown or reboot to
want to wait for all instances of a certain type to complete their work
before proceeding.

Services that implement networking are expected to order itself before
network.target on startup and hence after network.target on shutdown,
so that it should be sufficient to just have

After=network.target

_______________________________________________
systemd-devel mailing list
systemd-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/systemd-devel

_______________________________________________
systemd-devel mailing list
systemd-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/systemd-devel

Reply via email to