Re: edac_core: crashes on shutdown

[Doug Thompson <norsk5@yahoo.com>]

--- On Thu, 12/2/10, Borislav Petkov <bp@amd64.org> wrote:

> From: Borislav Petkov <bp@amd64.org>
> Subject: Re: edac_core: crashes on shutdown
> To: "Tobias Karnat" <tobias.karnat@googlemail.com>
> Cc: "linux-edac@vger.kernel.org" <linux-edac@vger.kernel.org>, "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>, "Doug Thompson" <norsk5@yahoo.com>
> Date: Thursday, December 2, 2010, 11:37 AM
> On Thu, Dec 02, 2010 at 01:05:01PM
> -0500, Tobias Karnat wrote:
> > This patch fixed it.
> > I have rebooted five times and it does not crash
> anymore.
> > 
> > Thank you.
> 
> Thanks for testing and taking pictures :). I'll send it to
> Linus before
> .37 is out.

great find. I haven't had time to do much with this one (sorry). Just started a new job (in user space now, instead of kernel space) in the past 2 weeks. As things settle down, I hope to have some cycles back on this.

doug t

> 
> > Btw, are there any information available regarding the
> NMI option?
> > 
> > parm:    edac_op_state:EDAC Error
> Reporting state: 0=Poll,1=NMI (int)
> > 
> > In edac.txt NMI is listed under FUTURE HARDWARE
> SCANNING.
> 
> Well, looking at
> <arch/x86/kernel/traps.c:mem_parity_error()> this
> should already work. But it is kinda of a hack, if I'm
> reading Doug
> correctly: http://lkml.org/lkml/2010/9/21/144
> 
> And yes, using some kind of an interrupt is much better
> than polling but
> I don't know whether there's a single interrupt source for
> the error
> types all edac drivers can decode and report. In the
> amd64_edac case,
> we're piggybacking on MCE, for example. This was actually
> the initial
> reason for 00740c58541b6087d78418cebca1fcb86dc6077d and
> dropping polling
> from that driver.
> 
> Doug, any additions?

I had an implementation for NMI processing, but as pointed out, we don't have a single cause source that applies everyone, and I (we) didn't have enough boxes to try it out on.  The documentation indicated a future feature with NMI, but it has faded over time and NMI patch (at sourceforge) is way stale.

I whole new R&D stage would need to be done again to recapture all the old information and new possible implementation changes that have occurred in the past 4 years (if any).  

The NMI/parity was a hack that was used to solved a problem back in 2005/2006 at Linux Networx (now defunct), but it worked then.  It was very much a lower priority thing once it worked on many machines. Today, it need a re-eval

doug thompson