[arch-general] mdadm segfaults with kernel 2.6.38.2

[Norbert Zeh]

Norbert Zeh [2011.04.11 2102 -0300]:
> Divan Santana [2011.04.11 2132 +0200]:
> > On Monday 11 April 2011 20:58:40 Norbert Zeh wrote:
> > > I just upgraded to the latest kernel.  The /etc/rc.d/mdadm script, which
> > > starts mdadm in monitor mode, reports a segmentation fault in
> > > mdadm --monitor --oneshot --scan.
> > > 
> > > I also tried to manually start, e.g., mdadm --monitor /dev/md0, and get
> > > a segmentation fault there, too.
> > > 
> > > Since this did not happen before the kernel upgrade and mdadm interacts
> > > rather closely with the kernel, I suspect that some change in the latest
> > > kernel is to blame for this.
> > > 
> > > Did anybody else run into the same problem?  
> > 
> > > Where would I start
> > > debugging it?
> > 
> > Downgrade to 2.6.37.5/6 and confirm it's the kernel.
> > Then search for a bug, join the bug report, or file one.
> 
> I had this on my to-do list already, but I appreciate the suggestion.
> Now, the good news is: it's not the kernel, as the same thing happens
> with the kernel I was running previously: 2.6.37.5.
> 
> The bad news: it still happens, and I think I know the culprit.  mdadm
> was upgraded along with my kernel upgrade.  So occam's razor suggests
> the new version of mdadm itself is to blame for the segfault.
> Unfortunately, while I diligently keep old versions of my kernel
> package, I don't have the previous version of mdadm lying around to
> investigate this further.  Any suggestions?

And things got even stranger just now, with a silver lining.  The
version of mdadm found in [core] is 3.2.1.  ABS, on the other hand,
still has the old version (3.1.5).  Isn't ABS supposed to be the
official source tree for the official repos?  Maybe I misunderstand the
relationship between ABS and the repos.

In any case, given that ABS still had the old version, I was able to
rebuild the old version of mdadm and downgrade.  Voila, it works, with
the newest kernel.  So the culprit is the new mdadm version.  I guess
it's time to report the bug upstream and, for now, stick to the old
version of mdadm.

Cheers,
Norbert