[Linux-cluster] DLM Problem

[gordan at bobich.net]

On Thu, 31 Jan 2008, isplist at logicore.net wrote:

>> You could just try it without LVM alltogether. LVM is entirely optional in
>> this.
>
> I didn't know that to be honest. I think you had mentioned it in another post.
> Since day one my understanding was that GFS had to be set up using LVM.
> What are the steps to creating a GFS volume without LVM?

Yes, all the documentation seems to be hammering LVM down user's throat 
for some unfathomable reason. You can just mkfs.gfs -myoptions 
/dev/mygfsblockdevice

Under myoptions you'll want to specify the number of journals and lock 
module as dlm as per the docs.

> PS: I cleanly rebooted everything again. The logs still show the following;
>
> //
> Jan 31 09:44:45 compdev kernel: dlm: could not bind to local address for
> connect: -98
> \\

That vaguely rings a bell. I seem to remember that the problem in my case 
was somewhere in cluster.conf disagreements between nodes, and package 
versions not being up to date. Also make sure you start various things 
using the scripts under /etc/init.d, rather than starting things up 
manually.

What distro are you using?

> And, I hate to post such long winded logs but it might be helpful. This is
> where things get weird. I've posted about this several times. There is an
> error message that keeps trying to show up as vgchange is trying to run.
>
> There seems to be old volume information somewhere on the system causing this
> and I think this is the root of the problem. If I could remove the old long
> dead information, I think I could get this resolved.

Just blow away the LVM (uninstall the clvm package, just to make sure),
dd if=/dev/zero > /dev/mygfsblockdevice bs=1024 count=1024
(just to make sure it is properly gone - assuming you don't have any 
important data on there, of course!)
and then re-create gfs on the block device as described above.

Start up cman, gfs, ccsd services, and it should work, if your 
cluster.conf is working. Also remember that you need nodes to be quorate 
before gfs will successfully complete mounting. You can try it with 
locking lock_nolock with just a single node, to make sure your GFS file 
system is OK first.

Gordan