On Wed, Feb 05 2020, Wols Lists wrote:

> On 05/02/20 11:51, Paul Dann wrote:
>> Hi there,
>> 
>> I've got a load of data on an md RAID 1 array assembled from 2x4TB
>> disks. I'm looking to expand this by adding a third 4TB disk and
>> converting the array to RAID 5. Now the required procedure is
>> documented on the wiki, but my question is:
>> 
>> When I convert the RAID 1 array to RAID 5, the array will be in a
>> degraded state as it rebuilds onto the new disk. However, if one of
>> the original two disks were to fail during this procedure, is mdadm
>> smart enough to convert the array back to degraded RAID 1, or will my
>> array now be a broken RAID 5 with no path to recovery?
>
> There is a "revert reshape" option which will take you back to a raid 1.
> This assumes, however, that it's the new disk that has failed.
>
> I'm pretty certain, however, that should one of the old disks fail
> during the conversion you will end up with a degraded raid 5.

This is correct.
It is not true to say "the array will be in a degraded state as it
rebuilds onto the new disk" as Paul did above.
During the rebuild the array will be in 2 parts:
 - a non-degraded raid5 which starts small and grows
 - a non-degraded raid1 which starts large and shrinks.

at no point in time is the array without full redundancy (unless a
device fails of course).

NeilBrown


>
> If you're worried, I would make sure you've done a SMART health check on
> your two original drives, although that's no guarantee everything's okay.
>
> If you really are that worried, get two new disks with the intention of
> ending up at raid 6. Add a 3rd drive to the mirror, fail and remove one
> of the originals, add the 4th to go raid 5, then add back the first to
> go raid 6.
>
> One thing I will say - MAKE SURE you have the latest mdadm, and if
> possible make sure you're running a recent kernel. There have been
> issues converting from 1 to 5. None of them serious, and all fixed by
> upgrading mdadm/kernel, hence the advice to be running the latest/greatest.
>> 
>> Many thanks,
>> Paul
>> 
> Cheers,
> Wol