Re: Questions about initialization process of mdadm

Re: Questions about initialization process of mdadm

[NeilBrown]

[-- Attachment #1: Type: text/plain, Size: 2253 bytes --]

On Tue, 4 Dec 2012 12:12:23 +0900 Taejin Kim <taejin1999@davinci.snu.ac.kr>
wrote:

> Dear Neil Brown,
> 
> I'm Taejin Kim, Ph.D student at Seoul National Universtiy in Korea.
> Recently, I got interested in RAID for flash memory and just started
> studying about RAID.
> 
> I have seen that there are considerable number of writes during
> initialization of mdadm, reconstructing array from degraded status.
> When I search for what mdadm does, I was able to find your article about
> the RAID5 initialization of mdadm so I now understand what happens.
> (http://marc.info/?l=linux-raid&m=112044009718483&w=2)
> 
> However, I'm wondering why we need to make sure the parity blocks are all
> correct even for a "new" RAID5 array.

Because when a single block is updated, the raid5 module  might use a
read-modify-write cycle.
i.e.
 - read old data and parity
 - subtract old data from parity, and add new data to parity
 - write new data and parity

If the parity block wasn't correct before, then it won't be correct after.

So we must make sure it is correct  at the start.

> In my point of view, there would be no meaningful data in a "new" drive so
> we can write data and parity block from the beginning without having
> consistent parity block calculated from meaningless data in a "new" drive.
> 
> I could come up with one possible reason for the initialization, which is a
> bad block management.
> Is it correct to recognize bad blocks in each drive during initialization
> by trying to write out parity blocks?
> If not, would you explain the reason for this initialization?
> 
> I'm also wondering if this process is necessary only for mdadm or for other
> RAID tools, too.

I believe most RAID implementations perform an initial sync.

md doesn't current require it for other levels (just RAID5 and RAID4) as they
don't do any read-modify-write. However it is possible that the
implementation for RAID6 might change  one day, so don't assume that
--assume-clean is safe for RAID6 long-term.


NeilBrown

> 
> I would appreciate if you let me know above questions. It would be very
> helpful for my research.
> I'm looking forward to your reply.
> Thank you.
> 
> Sincerely.
> 


[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 828 bytes --]