Re: RAID 5 rebuild fails with power interruption.

[Neil Brown <neilb@suse.de>]

On Tue, 24 Nov 2009 02:36:37 +0100
Kasper Sandberg <postmaster@metanurb.dk> wrote:

> On Tue, 2009-11-17 at 09:47 +1100, Neil Brown wrote:
> > On Mon, 16 Nov 2009 16:00:38 +0530
> > <senthilkumar.muthukalai@wipro.com> wrote:
> > 
> > > We face this problem in our NAS product where we handle RAID5.
> > > In the below mentioned scenario, when RAID5 is rebuilding after
> > > adding a disk, we reset the power.
> > > Ideally when the system comes up, the RAID5 should have accepted
> > > the disk but not in our case.
> > > We get the 'kicking the non-fresh disk from array' message with
> > > the boot message.
> > > In our RAID init script we run 'mdadm -- Examine -- scan',
> > > followed by 'mdadm --assemble'.
> > > Could you pls help me to understand why this disk is being thrown
> > > out?
> > 
> > It is because the metadata being used (v0.90) does not have the
> > ability to record that a device is partially recoverred.   It can
> > only record that a device is either a full member of the array, or
> > is not a member of the array.  So until the recovery completes, the
> > metadata only records that the device is not a member of the
> > array.  So when you restart, you find that the device is not a
> > member of the array.
> > 
> > 
> > > What could be the solution?
> > 
> > Use 1.x metadata.  e.g. add
> >    --metadata=1.1
> > to your --create command.
> > 1.x metadata is able to record that a device is only partially
> > recovered.  So when the array is restarted the device will be
> > included and recovery will continue.
> 
> Might it be possible to upgrade metadata without having to recreate
> the array?

It isn't currently possible.
It would not be too hard to implement a conversion from 0.90 to 1.0
metadata, but I have not concrete plans to provide this.

NeilBrown