[xfs-masters] Re: 2.6.18-rc6-mm2

[Andrew Morton <akpm@xxxxxxxx>]

On Fri, 15 Sep 2006 12:57:45 +1000
David Chinner <dgc@xxxxxxx> wrote:

> On Thu, Sep 14, 2006 at 07:01:38PM +0200, Michal Piotrowski wrote:
> > >>
> > >> I'll build system with gcc 3.4
> > >
> > >It's not a compiler issue.
> > >
> > >Binary search should solve this mystery.
> > 
> > I was wrong - it's in vanilla tree
> > (http://www.stardust.webpages.pl/files/mm/2.6.18-rc6-mm2/mm-dmesg1).
> > 
> > cat hunt | head -n 3
> > origin.patch
> > BAD
> > libata-ignore-cfa-signature-while-sanity-checking-an-atapi-device.patch
> 
> Not sure what this means....

"BAD" is a bisection point, as per
http://www.zip.com.au/~akpm/linux/patches/stuff/bisecting-mm-trees.txt.  So
just 2.6.18-rc6+origin.patch exhibits the failure.  That is mainline.

> > I can reproduce this bug with all CONFIG_DEBUG_*=y.
> > (only
> > CONFIG_DEBUG_SYNCHRO_TEST=m
> > CONFIG_RCU_TORTURE_TEST=m
> > as modules)
> 
> I notice you're running i386 with 4k stacks - I wonder if you're blowing the
> stack by running xfs on loopback. I've been testing on x86_64 and ia64
> which don't have those issues. Can you try with 8K stacks instead of
> 4k stacks?

hm, that wouldn't be good.

Enabling CONFIG_DEBUG_STACK_USAGE will make the fourth column in the
sysrq-T output display the minimum-ever free stack space for each task.

sleep         S ffff810100f0bea8     0 18893  22372                     (NOTLB)

                                     ^ this number.