Re: raid and sleeping bad sectors

[Mike Tran <mhtran@us.ibm.com>]

On Tue, 2004-06-29 at 15:56, Dieter Stueken wrote:
> Mike Tran wrote:
> > (Please note that I don't mean to advertise EVMS here :) just want to
> > mention that the functionality is available)
> > 
> > EVMS, (http://evms.sourceforge.net) provides a solution to this "bad
> > sectors" issue by having Bad Block Relocation (BBR) layer on the I/O
> > stack.
> 
> Before proposing any solutions, i think it is very important to
> distinguish carefully between different kinds of errors:
> 
> a) read errors: some alert bell should ring (syslog/mail..)
>     but the system should not careless disable any disk to avoid
>     making the problem even worse.
> 
> b) write errors: if some blocks are written partly, but can not
>     be written to all disks, it may help, to write the data
>     (may be temporary) somewhere else.
> 
> when we got a read error, due to an unreadable sector, we may
> first try to rewrite it. In most cases, bad sector replacement
> of the HD-firmware takes action and the problem is solved so far.
> 

For raid1 mirroring, I think the code for "rewrite" does not look too
bad.  For raid5/raid5, it's going be harder.  I'm not saying that it's
not doable :)

In fact, there is a cnt_corrected_read field in the MD ver 1
superblock.  So, I hope this feature is coming soon.


> Only after this failed, we should turn over to plan b)
> 
> case b) may also help, if some disk gets temporary unavailable
> (i.E. cabling problem). After manual intervention, that brings
> the disk back on line again, the redirected data may even be
> copied back.
> 

Plan b) needs that "somewhere else."  This can also be achieved with the
MD ver 1 superblock.  Where we can reserve some sectors by correcly
setting the usable data_offset and data_size.

Now, we need user-space tool to create MD arrays with ver 1 superblock. 
In addition, of course, we will also need to enhance MD kernel code.


Cheers,
Mike T.