Re: RAID reconstruction hangs the whole system

[Greg Oster <oster%cs.usask.ca@localhost>]

On Thu, 5 Dec 2013 23:30:06 +0200
Jarmo Jaakkola <netbsd-users%roskakori.fi@localhost> wrote:

> On Thu, Dec 05, 2013 at 02:04:09PM -0600, Greg Oster wrote:
> > On Thu, 5 Dec 2013 21:37:53 +0200
> > Jarmo Jaakkola <netbsd-users%roskakori.fi@localhost> wrote:
> > > Is [RAIDframe only supporting two mirrored components] still valid
> > > (NetBSD 6.1.2)? 
> >
> > Yes.
> >
> > > What would this actually mean?
> > It means you should only use 2 components for a RAID 1 set, as
> > using 3 or more is not currently supported.
> 
> Sorry, my meaning didn't come through properly.  I meant: If one were
> to not heed this advice, what should one expect to happen?

So what RAIDframe does internally is if you specify 4 components, say
A, B, C, and D, it arranges them like this:

  A  B
  C  D

so that A and B are mirrors of each other, C and D are mirrors of each
other, and A and C are concatenated together (just like B and D). 

If you check your disklabels and such, I think what you'll see is the
size of your RAID 1 set is actually about 2x what it should be.... 

> > > Should raidctl barf when trying to create a mirrored set with more
> > > than two components? 
> > 
> > It probably should, but I'm betting it doesn't.  What does your
> > RAID config file look like?  What does 'raidctl -s' say for the RAID
> > set?
> 
> The RAID set configured and seems to work just fine except for
> the  reconstruction problems.

I'm surprised it did, as it's technically missing the '4th component'
that it would need to work properly....   Essentially your wd4a is not
mirrored anywhere, and doesn't have anywhere to rebuild to -- and it
might be this later fact that is causing things to hiccup when you try
to rebuild.....  Havn't looked at all the relevant code, but at least
the RAID 1 config bits in the kernel don't seem to check to make sure
there are an even number of components provided, and I'm betting
raidctl doesn't do so either :(  (So ya it 'worked', but wasn't really
providing the RAID 1 like you were expecting :(  At a minimum I should
have added more error checking to enforce the even number of components
requirement...

Later...

Greg Oster