Re: 2.6.17-rc5-mm1

[Andrew Morton <akpm@osdl.org>]

On Sun, 4 Jun 2006 20:59:50 -0400
Chuck Ebbert <76306.1226@compuserve.com> wrote:

> In-Reply-To: <44833955.9000104@free.fr>
> 
> On Sun, 04 Jun 2006 21:49:41 +0200, Laurent Riffard wrote:
> 
> > > Something strange is happening with the stack.  Can you try with 8K stacks?
> > > 
> > > kernel hacking --->
> > >    [ ] Use 4Kb for kernel stacks instead of 8Kb
> > > 
> > 
> > Good catch!
> 
> Jan Beulich was the one who noticed the stack overflow.
> 
> > I just tried with 2.6.17-rc5-mm3. The BUG still happens with 4K stacks,
> > but the system runs fine with 8K stacks.
> 
> Can you try -mm3 with "check for stack overflows" and 4k stacks?

Maybe we could modify the check-for-stack-overflows code with 8k stacks to
be "check for overflows which would kill a 4k stack kernel".  By decreasing
the thresholds to 3k (say).

> > Another info: the system is able to run fine with the following scenario,
> > even with 4K stack:
> > - boot to runlevel 1
> > - load pktcdvd (modprobe + pktsetup)
> > - then go to runlevel 5 
> > It fails if pktcdvd is loaded at runlevel 2 or higher.
> 
> I have no idea why that would happen.

I'd be surprised if it has to do with the packet code - more likely someone
screwed something up in the lockdep/genirq/unwind area and the kernel went
recursive.

I tried it here, couldn't reproduce it.  Laurent, can you (re?)send the
offending .config?