Re: [patch] ext2/3: document conditions when reliable operation is possible

[Rik van Riel <riel@redhat.com>]

Pavel Machek wrote:

> Ok, can you help? Having a piece of MD documentation explaining the
> "powerfail nukes entire stripe" and how current filesystems do not
> deal with that would be nice, along with description when exactly that
> happens.

Except of course for the inconvenient detail that a power
failure on a degraded RAID 5 array does *NOT* nuke the
entire stripe.

A 5-disk RAID 5 array will have 4 data blocks and 1 parity
block in each stripe.  A degraded array will have either
4 data blocks or 3 data blocks and 1 parity block in the
stripe.

If we are dealing with a parity-less stripe, we cannot
lose any data due to RAID 5, because each of the 4 data
blocks has a disk block available.  We could still lose
a data write due to a power failure, but this could also
happen with the RAID 5 array still intact.

If we are dealing with a 3-data, 1-parity stripe, then
3 of the 4 data blocks have an available disk block and
will not be lost (if they make it to disk).  The only
block that maintains on all 3 data blocks and the parity
block being correct is the block that does not currently
have a disk to be written to.

In short, if a stripe is not written completely on a
degraded RAID 5 array, you can lose:
1) the blocks that were not written (duh)
2) the block that doesn't have a disk

The first part of this loss is also true in a non-degraded
RAID 5 array.  The fact that the array is degraded really
does not add much additional data loss here and you certainly
will not lose the entire stripe like you suggest.

-- 
All rights reversed.