Re: Online resize

[Michael Tokarev <mjt@tls.msk.ru>]

Wayne Gemmell wrote:
> On Wednesday 12 March 2008 10:04:56 you wrote:
>> a) re-create the array with new partitions, with THE SAME DISK ORDER,
>>    METADATA VERSION AND CHUNK SIZE as it was before.
> Is this possible without loosing data?

Yes.  But it's unreliable, sorta.  Too easy to screw things up, that is.

>> b) restore partitions to the EXACT size as they were before, and
>>    start from scratch
> Thats easy enough. I'll do that now.

And now, the original question comes up again: what to do, how to
perform the actual resize of the raid array...

To be fair, I don't really know.  I always used my own tiny program
to assist me in a situation like this - to back up and restore
v0.90 linux raid superblocks.  It's available at
http://www.corpit.ru/mjt/mdsuper.c .  Here's how I use it:

First, save the original superblocks in files:

  for f in sd{a,b,c,d,e}1; do
    ./mdsuper read /dev/$f > $f
  done

this instructs mdsuper to read the superblock (v0.90 only!) and
write it to standard out.  This has to be done while the array
(on original partitions) is STOPPED.

Now, resize the partitions, reboot if necessary (for the kernel
to recognize the new sizes).  You can save a dump of your current
partitions using `sfdisk -d' (redirect its output to a file as
with mdsuper) - just in case if something goes wrong.

At this point, your array is unaccessible - mdadm will tell you
that it can't find the superblocks.  But we have the superblocks
saved - so we only need to restore them back:

  for f in sd{a,b,c,d,e}1; do
    ./mdsuper write /dev/$f < $f
  done

(note the redirection is opposite!, and read was changed to write!).

Now, try to assemble the array - it should be here.  We just moved
(actually copied) the superblock from the now-middle-of-partitions
to the end, where it should be, and mdadm is now able to find it.
The array is still of the old size - but now mdadm --grow will
work.

That's basically it.  Just watch the mdsuper invocation.

/mjt