On 07.12.2015 13:58, Daniel P. Berrange wrote:
On Mon, Dec 07, 2015 at 01:57:20PM +0100, Michal Privoznik wrote:
On 07.12.2015 13:55, Daniel P. Berrange wrote:
When installing the libvirt-daemon RPM, we have a %post rule to enable the libvirtd.service, virtlockd.socket and virtlogd.socket files. This is only done, however, when the RPM is first installed, not when upgrading RPMs.
This adds a separate rule to explicitly enable the virtlockd.socket and virtlogd.socket units upon RPM upgrade, if libvirtd.service is marked as enabled. It also adds a rule to start the socket units if libvirtd.service is running.
The downside with taking this approach is that it is impossible for a user to have libvirtd.service enabled and disable socket activation, because we'll re-enable it on every upgrade. Similarly we'll be starting the socket listeners on every upgrade if libvirtd is running. This is a pretty nasty thing to have to keep in libvirt for the long term, given this upgrade problem is a one-time only pain point.
For this reason I am recommending *NOT* applying this patch to GIT.
Instead, I think we just have to release note the fact that after upgrading you should enable the virtlockd.socket unit.
--- libvirt.spec.in | 12 ++++++++++++ 1 file changed, 12 insertions(+)
diff --git a/libvirt.spec.in b/libvirt.spec.in index 1f76f24..48e5632 100644 --- a/libvirt.spec.in +++ b/libvirt.spec.in @@ -1703,6 +1703,18 @@ if [ $1 -eq 1 ] ; then libvirtd.service >/dev/null 2>&1 || : fi %endif + +# In upgrade scenario we must explicitly enable virtlockd/virtlogd +# sockets, if libvirtd is already enabled and start them if +# libvirtd is running, otherwise you'll get failures to start +# guests +if [ $1 -ge 1 ] ; then + /bin/systemctl is-enabled libvirtd.service 1>/dev/null 2>&1 && + /bin/systemctl enable virtlockd.socket virtlogd.socket + /bin/systemctl is-active libvirtd.service 1>/dev/null 2>&1 && + /bin/systemctl start virtlockd.socket virtlogd.socket +fi + %else %if %{with_cgconfig} # Starting with Fedora 16/RHEL-7, systemd automounts all cgroups,
ACK
Presumably you didn't read my commit message, or are you ACK'ing my suggestion that we *NOT* merge this ?
Well, I think that socket activation is nice feature. But as even DV pointed out (and I've experienced that too on my non-systemd system) that it's impossible to start domains without virtlogd enabled. I find it much worse than depending on a 'nice systemd feature'. Therefore I think we should merge it. Michal