Bug 496857

Created attachment 340524 [details]
This patch implements the /etc/xtab approach.

So, you are mounting network filesystem A, running things from it, and then mounting root filesystem B?

That's... a little pathological.

I agree, that you won't need this for locally used filesystems that are not dependent on userspace processes.

Nevertheless, there are basically two usecases that require attention during the halt process.

usecase_1: Filesystems that depend on userspace processes. (e.g. NFS4, GFS, GFS2)

usecase_1 requires a chroot environment for the userspace processes that are needed to mount the root filesystem. E.g. GFS needs the cluster stack up and running (aisexec, clvmd, ccsd, ...) 

During the halt process, the chroot environment must not be umounted before the root filesystem. To be able to umount the root filesystem, a pivot_root to the chroot environment has to be done.

proposed halt process (right now implemented in /sbin/halt.local which is implicitly called by /etc/init.d/halt):

1. cd <chroot_environment> 
2. pivot_root . /mnt/oldroot
3. init u (releasing the old root filesystem)
4. umount /mnt/oldroot
5. leave cluster and cleanup
6. reset or halt

if the chroot environment is umounted before the root filesystem, the node either stalls or the cluster needs to initiate an unneeded and time consuming filesystem recovery.

usecase_2: Shared root configuration with hostdependent files.

usecase_2 requires a --bind mount of a hostdependent dependent area in the root filesystem. I.e. we bind mount a special directory residing on the root filesystem to another place (/cluster/cdsl/<nodeid> to /cdsl.local).

e.g. mount --bind /cluster/cdsl/1 /cdsl.local

Then /var which has to be hostdependent points to /cdsl.local/var. 

Right now the halt initscript umounts this --bind mount and therefore the halt cannot proceed. This leads to an unsuccessful umount of the root filesystem causing the cluster to initiate an unneeded and time consuming filesystem recovery.

In my opinion, for the described use cases it is crucial to leave some dependent filesystems mounted before unmounting the root filesystem. 

I hope this makes it clear why it would be good to have this flexibility in the initscripts.

(In reply to comment #3)

> usecase_2: Shared root configuration with hostdependent files.
> 
> usecase_2 requires a --bind mount of a hostdependent dependent area in the root
> filesystem. I.e. we bind mount a special directory residing on the root
> filesystem to another place (/cluster/cdsl/<nodeid> to /cdsl.local).
> 
> e.g. mount --bind /cluster/cdsl/1 /cdsl.local
> 
> Then /var which has to be hostdependent points to /cdsl.local/var. 

This case does not seem different to me than network root; in this
case, the filesystem that is bindmounted should be marked in the exact same way a network root filesystem is.

The _netdev mountoption does not find its way into /proc/mounts and in /etc/init.d/halt it is not checked:

LANG=C __umount_loop '$2 ~ /^\/$|^\/proc|^\/dev/{next}
        $3 == "tmpfs" || $3 == "proc" {print $2 ; next}
        /(loopfs|autofs|nfs|cifs|smbfs|ncpfs|sysfs|^none|^\/dev\/ram|^\/dev\/root$)/ {next}
        {print $2}' /proc/mounts \
        $"Unmounting file systems: " \
        $"Unmounting file systems (retry): " \
        -f

nfs is explicitly excluded but this will only work for nfs not for any other cluster filesystem.

Or did I mix or miss understand something?

Marc.

Initscripts no longer handle this -> close.