There's some problems with the initscript dependencies between the various components that make up a working gfs cluster. F.e. cman provides cluster-manager but is referenced as cman in the dlm initscript I've checked them all and changed the provides and need clauses. Everything is referencing cluster-<package> now. Attached are diffs against the .rc files.
Created attachment 100824 [details, diff] dependencies ccsd
Created attachment 100825 [details, diff] dependencies clvmd
Created attachment 100826 [details, diff] dependencies cman
Created attachment 100827 [details, diff] dependencies dlm
Created attachment 100828 [details, diff] fenced deps
Created attachment 100829 [details, diff] deps gfs
Created attachment 100830 [details, diff] deps rgmanager
Could someone with a little more expertise take a look at these patches? Note that the patches are reversed. Ramon, it's great that you're providing patches. In the future, could you create them in the order of original first, then modified file? This will reduce confusion and allow easy application. Does the current situation actually cause any problems? It's just not yet abstracted as nicely as one would like, AFAICT.
Oops sorry for the reversed patches. Yes it creates problems. Ideally you would like to add gfs initscript to a runlevel and have the dependency system figure out what to do with the dependencies. And in which order to start them. Right now that doesn't work correctly. IIRC the locking-manager doesn't get started or too late in the process. Additionally there are problems with starting clvm, in some cases it needs network connectivity and/or device discovery before it can operate correctly. This isn't present at all in the initscripts and probably needs to be abstracted out into the conf.d/scripts. It depends on the type of shared storage hardware you're using. I have a working solution for my environment, but haven't had the time to move it into conf.d. setup. Right now it's a rather ugly hack on the initscript to poll for device presence.
dlm init script has a wrong reference to cman, but I'm not completely sure that renaming ccs, fence, etc provide/depend is a good thing. Said the last period was very bad for me and I haven't completed the clvm init scripts rewrite. I plan to make two big changes: - start only local non clustered lvm volumes in /lib/rcscripts/addons/lvm-* scripts - start clustered volumes after ccs, cman, dlm, fence, clvm imho this is the only way to make this working as expected (obviously we cannot automount gfs volumes before starting the locking manager, cluster manager and so on).
Given the age of this ticket and the fact that a few packages mentioned aren't in the tree anymore.. Is there any chance of resolving this?
I'd say no, but I've also stopped using gfs behind due to performance problems with it, maybe there are still peeps that actually use it ? The bug is about gfs1 which is superseded by gfs2 AFAIAC close it
(In reply to comment #12) > AFAIAC close it Agreed