Re: [BUG] x86: bootmem broken on SGI UV

[Robin Holt <holt@sgi.com>]

On Sun, Oct 10, 2010 at 04:05:54PM +0200, Ingo Molnar wrote:
> 
> * Robin Holt <holt@sgi.com> wrote:
> 
> > On Sun, Oct 10, 2010 at 06:44:59AM -0500, Robin Holt wrote:
> > > On Sun, Oct 10, 2010 at 12:43:19PM +0200, Ingo Molnar wrote:
> > > > 
> > > > * Ingo Molnar <mingo@elte.hu> wrote:
> > > > 
> > > > > 
> > > > > * Robin Holt <holt@sgi.com> wrote:
> > > > > 
> > > > > > On Sat, Oct 09, 2010 at 07:59:45AM -0500, Russ Anderson wrote:
> > > > > > > Yes, Yinghai's patch fixes the problem.
> > > > > > > Thank you very much.
> > > > > > 
> > > > > > Will this be included in 2.6.36?  It is needed for boot in order for UV
> > > > > > systems to boot.
> > > > > 
> > > > > -tip uses memblock APIs. If this happens with vanilla -git as well 
> > > > > then we need a bootmem backport for the fix.
> > > > 
> > > > And to answer your question: yes, we can queue it up for -final as well 
> > > > if it's a recent regression - 'doesnt boot at all' bugs are nasty. But 
> > > > i'm not sure this is a bootmem problem so please double check vanilla 
> > > > v2.6.36-rc7 as well.
> > > 
> > > The 36-rc7 kernel does not boot at all either.  I don't have any 
> > > decent debug tools to dig in further.  It does fail on the same 
> > > machine that Russ was testing with, but passes on any that have a 
> > > single blade as the kernel Russ first identified as being a problem 
> > > had.  Based upon my vague recollection of the boot messages, it 
> > > appears to fail in a similar point in boot.  I would assume it is a 
> > > similar problem.
> > 
> > 2.6.35 fails as well.  I will bisect for as long as time permits.
> 
> That's really bad. If this means that you have not booted vanilla 
> mainline on UV in that timeframe _at all_, and that it was perma-broken 
> since 2009 when that commit Yinghai identified went upstream, then 
> there's little point in squeezing this fix into v2.6.36-final.
> 
> -rcs are strictly for regression fixes.
> 
> If you want upstream to care about you then you absolutely have the duty 
> to at minimum test latest mainline ...

Upstream boots on many UV systems.  Some have more constrained memory
and end up with this weird mapping which causes the problem.  I am not
sure how many of the machines end up with this config, but it is far
from all of them.

Jack does build and boot a kernel every night on many different
configurations, just not this one.  I believe he and Russ are working
on getting this configuration into his nightly testing.

Robin