Re: RAID grow and disk failure

[Neil Brown <neilb@suse.de>]

On Sat, 26 Jun 2010 15:12:35 +0200
Piergiorgio Sartor <piergiorgio.sartor@nexgo.de> wrote:

> Hi,
> 
> > Assuming the code doesn't have any bugs, the reshape will stop, then
> > immediately restart picking up where it left off.
> 
> thanks, that's what I wanted to know.
> 
> > You will of course end up with a degraded array
> 
> Yes, that was clear.
> 
> > It might be nice in these circumstances to abort the reshape and revert back
> > the the previous number of devices - particularly if it was the new device
> > that failed.  However that currently isn't supported.
> 
> Well, probably as an option, it could be interesting.
> 
> Actually, I would be still interested, we already
> discussed the topic, on a RAID-5/6 with HDDs of
> different size.
> This would simplify many things...
> 
> > > 1)
> > > mdadm --grow ...
> > > mdadm --wait
> > > pvresize
> > 
> > Yes.
> > 
> > > 
> > > 2)
> > > mdadm --grow
> > > pvresize
> > 
> > No.
> > Until the reshape has completed, the extra space is not available.
> 
> There seem to be an issue, here, maybe.
> 
> Using the command line:
> 
> mdadm --grow /dev/md/vol02 --bitmap=none; mdadm --grow /dev/md/vol02 -n 9 --backup-file=/var/tmp/md125.backup; mdadm --wait /dev/md/vol02; mdadm --grow /dev/md/vol02 --bitmap=internal --bitmap-chunk=128
> 
> Note that /dev/md/vol02 is the usual link to /dev/md125,
> which should be the same for this scope, I guess.
> 
> I got (in two independent tests):
> 
> mdadm: Need to backup 2688K of critical section..
> mdadm: failed to set internal bitmap.
> 
> Re-issuing:
> 
> mdadm --wait /dev/md/vol02; mdadm --grow /dev/md/vol02 --bitmap=internal --bitmap-chunk=128
> 
> Does wait.
> 
> Could it be the devices (being USB) are so slow
> that some race condition is uncovered and the
> immediate "--wait" after the "--grow" does not work?
>

Yes, there  is a race here.  The reshape doesn't quite start instantly and so
--wait doesn't notice.
I've added a note to my todo-list to look into this.
For now, a 'sleep 1' between the --grow and the --wait should be enough.

Thanks,
NeilBrown