svn commit: r186038 - head/sbin/geom/misc

[Ulf Lilleengen]

On Sun, Dec 14, 2008 at 08:21:01PM +0100, Pawel Jakub Dawidek wrote:
> On Sat, Dec 13, 2008 at 02:14:56PM +0000, Ulf Lilleengen wrote:
> > Author: lulf
> > Date: Sat Dec 13 14:14:56 2008
> > New Revision: 186038
> > URL: http://svn.freebsd.org/changeset/base/186038
> > 
> > Log:
> >   - When writing metadata to a geom provider, open the it as read-write since it
> >     might do subsequent reads from other providers. This stopped geli (and
> >     probably other classes using g_metadata_store as well) from being put on top
> >     of gvinum raid5 volumes.
> >   
> >   Note:
> >   The reason it fails in the gvinum raid5 case is that gvinum will read back the
> >   old parity stripe before calculating the new parity stripe to be written out
> >   again.  The write will then fail because the underlying disk to be read is
> >   opened write only.
> 
> I think we discussed this in the past. In my opinion this change is
> incorrect. The intend here is to only write something, that's why I use
> O_WRONLY. If gvinum/raid5 needs also read something to satisfy the
> write, this is gvinum/raid5's internal detail and should be handled
> there. RAID5 by design needs to read before it can write (in most
> cases), does it mean that the O_WRONLY flag became bogus suddenly? No.
> The flag given to open(2) describes caller's intend, not provider's
> internals.
> 
Hmm, I understand your reasoning. But how can gvinum read from these internal
providers then? Is there some way to override this internally in GEOM?

-- 
Ulf Lilleengen