[ale] The perpetual question: best current HDD?

[Phil Turmel]

On 01/10/2013 12:31 PM, Derek Atkins wrote:

> I *assume* that the weekly raid-check (which sends "check" into
> /sys/block/$dev/md/sync_action) is a read-write scrub.  This would be
> used, as you say, to find weak sectors and remap them on the drives.

Not quite.  It is a read scan of all data, mirror, and parity blocks,
but the normal rewrite logic is triggered if a read error is encountered.

In general, drives don't read back every sector they write to make sure
it was written correctly, as that would dramatically slow write speed.
So it is not safe to rewrite the same data, as that can silently corrupt
it, especially when the drive is beginning to wearing out.  MD only
rewrites a sector from mirror/parity if it cannot read it.

If you use a "repair" scrub, MD will unconditionally write the first
mirror's data to all others, or unconditionally compute and write fresh
parity blocks.  It will still reconstruct on a read error as usual.

There's been some traffic on linux-raid about an implementation of
"repair" that only writes if there's a mismatch, but I don't think it
has been merged.

Phil