On Fri, 05 Dec 2014 12:27:31 +0300 Michael Tokarev [off-list ref] wrote:

Hello.

We in Debian have 2 rather similar bugreports which are somewhat old:
 http://bugs.debian.org/599352
 http://bugs.debian.org/694513
Last bugreport is against mdadm-3.2.5, so not THAT old.

Basically, it smells like there's a possibility for mdadm to assemble
more than one (degraded) array out of components of the same array.

So I wonder if mdadm or kernel have some protection of this situation
from happening.  Something like looking at already active/assembled
devices to see if the same UUID is already used before trying to
assemble another array.

It might be better done in kernel because this way the process might
be freee from races when two mdadm instances tries to do the work in
parallel.

Are these bugs for real?

Certainly are.
Fix in 3.3 I think.

commit 0431869cec4c673309d9aa30a2df4b778bc0bd24
Author: NeilBrown [off-list ref]
Date:   Wed Oct 10 18:27:32 2012 +1100

    Fix up interactions between --assemble and --incremental

is one part of the fix.  I cannot quickly find another patch which obviously
relates to the symptoms described, but my vague memories suggest there was
more than one patch.

NeilBrown