Re: [OSR-devel] iSCSI Root

[Gordan B. <go...@bo...>]

On Fri, 12 Oct 2007, Marc Grimme wrote:

>>>> It throws up a worrying error when it boots:
>>>> GFS: fsid=cluster:root.0: warning: assertion
>>>> "gfs_glock_is_locked_by_me(ip->i_gl)" failed
>>>> GFS: fsid=cluster:root.0 function = gfs_readpage
>>>> GFS: fsid=cluster:root.0 file =
>>>> /builddir/build/BUILD/gfs-kmod-0.1.16/_kmod_build_/src/gfs/ops_address.c
>>>> , line 279
>>>> GFS: fsid=cluster:root.0: time = 1192119131
>>>>
>>>> I wonder if this may be caused by a file system perhaps not having been
>>>> cleanly unmounted on a previous try while I was building it...
>>>
>>> Perhaps just fschk it when you're in the initrd.
>>
>> Yup, just did. Some minor things were broken with the fs metadata. But
>> when I reboot, I still get a similar message when HALd loads. I wonder if
>> I can safely switch that off - assuming that's causing it...
>
> You could also file it to the gfs list and see what they'll tell you. It seems
> to me I've seen this message also.

If you mean the RedHat's Linux Cluster list - I already did. :-)

>> Now, in theory, I should be able to bring up another node on the same file
>> system. All I would need to do is clone the /boot partition to the other
>> box, and it should just come up.
>
> Why cloning it and not using the same. Isn't that possible. We are always
> doing it this way.

Because I'm not booting this off DHCP. I'm booting the kernel and the 
initrd off the local disk. So I need to clone the boot partition with the 
kernel and the initrd to each of the nodes.

>> What do I need to do to achieve this, and can it all be done with the one
>> node that is already running? I'm assuming that I'll have to do something
>> like:
>>
>> mount --bind /cluster/cdsl/4/ /cdsl.local/
>
> exactly if nodeid is 4. But again the initrd should do this job automatically.

So, I wouldn't need to do this at all? The initrd will automagically link 
/cdsl.local to /cluster/cdsl/nodeid ?

>> As far as unsharing things under /var, I _think_ only /var/lock actually
>> needs to be unshared. Can I do this with the running image with:
>>
>> com-mkcdsl -r / -a /var/lock
> you can skip the -r/ it is default.
> How about /var/run, /var/log, /var/cache, /var/tmp, /var/spool. All of these
> normally need to be hostdependent.

I'm not sure why /var/cache and /var/spool would need to be host 
dependent. I can see reasons why I'd want to them to be shared.

I agree that /var/run and /var/lock should be private.

It would be _nice_ to have a shared /var/log, but from past experience, 
the logs will get messed up when multiple syslogs try to write to them. 
Is there a shared logging solution for this? I know I can pick a master 
log node and get syslog pointed at this, but this won't work for all the 
other non-syslog services (e.g. Apache).

I plan to link /var/tmp to /tmp, and have /tmp mounted to a big local 
partition (local disks are only planned to have /boot, /tmp and swap).

Which brings me to the next question - how do I use a local disk partition 
instead of the initrd? What's the procedure for that? It seems a more 
efficient solution than relying on a ramdisk that eats memory after 
booting up when there is plenty of local disk space available. How do I 
use /etc/sysconfig/comoonics-chroot ?

Gordan