Am 12.01.2014 13:11, schrieb Laurent Bigonville: > Le Sun, 12 Jan 2014 02:18:32 +0100, > Michael Biebl <bi...@debian.org> a écrit : > >> Am 07.01.2014 13:12, schrieb Laurent Bigonville: >>> Package: network-manager >>> Version: 0.9.8.8-2 >>> Severity: serious >>> >>> Hi, >>> >>> I'm opening this bug so this is not getting lost. >>> >>> When upgrading from the previous versions the following events might >>> occurs when the user is running systemd: >>> >>> 1) The package was not enabling the systemd unit in the previous >>> versions. >>> 2) NM was then started using the LSB script called network-manager. >>> This means that the cgroup was called network-manager.service. >>> 3) During the upgrade the systemd unit is properly enabled and the >>> LSB script is masked by a symlink, this means that at some point >>> during the upgrade, the canonical name changes from >>> network-manager.service to NetworkManager.service >>> 4) At the end of the upgrade, systemd looks for a >>> NetworkManager.service cgroups which is not existing and then >>> concludes that the service is not running at all. >>> 5) Systemd tries to start the service, NM complains that there is an >>> other instance of itself running => boom >>> >>> To fix this, I would propose to stop NM in the preinst script if we >>> are upgrading from the previous versions and that the user is >>> running systemd. >> >> Why preinst and not postinst (before the #DEBHELPER# stanza, i.e. >> before the service is enabled)? >> Stopping the service in preinst creates unnecessary downtime which >> could be fatal if NM is in charge of the network connection and you >> do a remote upgrade. >> I don't see a downside moving the network-manager stop to postinst, do >> you? Just in case I'm missing something > > Well, in the postinst the symlink that mask the LSB will already be > present on disk. That means that the canonical name of the service has > already changed. A daemon-reload is actually needed for systemd to > notice this change, but nothing guarantee that no other packages would > do it for us between the moment the files are unpacked and the moment > our postinst script is called.
Ah, k. I missed the part about the symlink/alias now being shipped directly in the package and it no longer being created via [Install] Alias=network-manager.service > That's why I thought that stopping the service in the preinst was the > safer choice. This somehow smells like the same bug as [1] or at least related, i.e. systemd getting confused when we mess around with the symlinks. (In case of rsyslog we basically go the other direction: the symlinks are no longer shipped in the package but created on install time via dh-systemd) I wonder if there is a way to address this directly in systemd. Michael [1] http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=724796 -- Why is it that all of the instruments seeking intelligent life in the universe are pointed away from Earth?
signature.asc
Description: OpenPGP digital signature
