Re: PROBLEM: OOPSes in PREEMPT SMP for AMD Opteron Dual-Core with Memhole Mapping (non tainted kernel)

[Jacob Martin <martin@cs.uga.edu>]

On Monday 13 June 2005 10:06 am, Andi Kleen wrote:
> On Sun, Jun 12, 2005 at 03:29:50PM -0400, Jacob Martin wrote:
> > Hardware memhole mapping never seems to work, or causes lockups right
> > away.  I need to test it further though.
> >
> > I have discovered that with the following features enabled:
> >
> > 1.  Software memhole mapping
> > 2.  Continuous,
> >
> > linux sees the entire 4GB of memory.  However, when things start getting
> > requested from the upper half, there are Oopses generated.  Attached are
> > two Oopses that occurred under the test scenario described.
>
> What happens when you boot with numa=off or with numa=noacpi ?

You got it!  It seems to be working just fine without it compiled into the 
kernel.

> The system seems to believe it has memory in an area not covered
> by mem_map.

I think you hit it right on the head.

I enabled NUMA because I had anticipated upgrading later.  So I guess if you 
don't actually have NUMA set up hardware-wise, and enable this module, then 
you will have problems.

Maybe a simple update to the kernel "K8 NUMA support" Processor feature's help 
section should be made to note this?  Or, is there something that could be 
fixed somewhere.  I wouldn't mind helping, it was baffling me for two weeks.

> > launch big memory apps.
> >
> > I suppose I could write a program to consume/probe the upper memory half.
> > Anyone know of a good/quicky way to do that?
>
> You can use the attached program which I often use for similar purposes.
> It writes nearly all free memory in a loop and also often triggers memory
> problems.

Thanks for that too!

Sincerely,
Jacob Martin


-- 
Jacob Martin