On Wed, 30 Jul 2014 20:24:30 +0000 Markus Stockhausen
[off-list ref] wrote:

Hi,

the last days I tried to understand the RAID6 logic when recalculating
P/Q parity (or syndrome) if only parts of a stripe are updated. As far
as I get the current implementation right it will always read the whole
stripe and build parities from scratch. E.g. updating a single chunk
on a 10 disk RAID6 will read the other 7 data blocks to write the block
plus 2 P/Q parities afterwards. Compared to the best case 3 read plus
3 write operations this can be quite a heavy impact. RAID5 at least
has a shortcut for that.

Being no specialist at all - but having some ideas - I would like to
get some feedback about the following way to improve the scenario.

1) Write a modified gen_syndrome function (call it xor_syndrome)
like in lib/raid6/xxx.c that allows to XOR an existing P/Q parity
without overwriting it. The cost of that function would be the same
as it only changes a few assembler commands.

2) Enhance the ops_run_prexor to call this xor_syndrome function
for an RAID6. Fill the not relevant pages (R5_wantdrain) with an
empty page or NULL. So the gen_syndrome will only find zeroes.
With this call P/Q will have a pre-xored version.

3) Enhance ops_run_reconstruct6 to allow processing of changed
blocks only. We will call xor_snydrome from there too, to XOR
the prexored P/Q once again.

4) Enhance handle_stripe_dirtying so it will allow calculation of
RCW versus RMW costs for the RAID6 case.

Did I miss something? And if not: Will the doubled CPU consumption 
for P/Q calculation justify that way of implementation? At least
we can avoid IOs on slow disks.

Maybe it has been discussed before but I did not find it on the
mailing list.

Thanks in advance.

Markus

Please see
  http://comments.gmane.org/gmane.linux.raid/42559

Yes, this is something we probably want.
The previous effort stalled somehow.  Maybe it just needs someone to start
pushing again.

NeilBrown