On 11/10/2013 4:54 PM, Adam Goryachev wrote:

On 11/11/13 09:36, Stan Hoeppner wrote:

quoted

On 11/10/2013 2:34 PM, NeilBrown wrote:

quoted

The firmware can only relocate a sector if it reads it when it is
marginal
but not yet completely lost.  If a sector is not read for a long time
and
during that time the media degraded beyond recovery the firmware
cannot do
anything.  But RAID1 can - it can get it from the other device.

But is a scrub required for this?  Isn't this exactly what occurs during
normal operation with md/RAID1?  I.e. a read fails with disk error, so
we grab the sector from the mirror?  So what advantage is there to
scrubbing md/RAID1?

Wouldn't a check of the raid cause each member to be read in full,
therefore helping the disk to notice that the sector is marginal, and/or
the RAID layer to notice that the sector is no longer readable and
therefore read from the other member, and re-write the sector. Consider
a sector that is very rarely accessed...

Or are you suggesting that a smart command issued to the underlying
devices can solve both of those scenarios?

No, what I suggest is that drive instrumentation will often alert one to
drive problems before you see a read error at the kernel.  Assuming this
is true then scrubbing isn't necessary.

What Neil describes is a case where a sector is written once and read
very infrequently, or possibly years after the write, i.e. long term
archiving.  In this case a scrub may discover a media defect which may
go unnoticed by the drive firmware or normal md array operation.

-- 
Stan