Re: ftp server crashes on heavy load: possible scheduler bug

["Pedro Venda (SYSADM)" <pjvenda@rnl.ist.utl.pt>]

On Friday 29 April 2005 15:21, Pedro Venda (SYSADM) wrote:
> On Friday 29 April 2005 13:08, Andrew Morton wrote:
> > "Pedro Venda (SYSADM)" <pjvenda@rnl.ist.utl.pt> wrote:
> > > We've made some changes on our ftp server, and since that it's been
> > > crashing frequently (everyday) with a kernel panic.
> > >
> > >  We've configured the 5 IDE 160GB drives into md raid5 arrays with LVM
> > > on top of that. All filesystems are reiserfs. The other change we made
> > > to the server was changing from a patched 2.6.10-ac12 kernel into a
> > > newer 2.6.11.7.
> > >
> > >  Not being able to see the whole stacktrace on screen, we've started a
> > >  netconsole to investigate. Started the server and loaded it pretty bad
> > > with rsyncs and such... until it crashed after just 20 minutes.
> > >
> > >  The netconsole log was surprising - "kernel BUG at
> > > kernel/sched.c:2634!"
> >
> > Strange.  It'd be interesting to try disabling CONFIG_4KSTACKS.  Also,
> > please add this to get a bit more info.
>
> hi,
>
> I'll try that. Should I do it with or without preemption?

kernel does not boot. I get (initite) repeats of 
"=======================================" (not scaled).

I'm testing with a preemptible kernel without 4kstacks but without the 
suggested patch.

regards,
pedro venda.

>
> regards,
> pedro venda.
>
> > diff -puN kernel/sched.c~a kernel/sched.c
> > --- 25/kernel/sched.c~a	2005-04-29 05:05:24.792004408 -0700
> > +++ 25-akpm/kernel/sched.c	2005-04-29 05:06:36.015176840 -0700
> > @@ -2631,7 +2631,12 @@ void fastcall add_preempt_count(int val)
> >  	/*
> >  	 * Underflow?
> >  	 */
> > -	BUG_ON(((int)preempt_count() < 0));
> > +	if ((int)preempt_count() < 0) {
> > +		printk("preempt_count=%d\n", preempt_count());
> > +		BUG();
> > +	}
> > +	if ((int)preempt_count() > 1000)
> > +		printk("preempt_count=%d\n", preempt_count());
> >  	preempt_count() += val;
> >  	/*
> >  	 * Spinlock count overflowing soon?
> > _

-- 

Pedro João Lopes Venda
email: pjvenda < at > rnl.ist.utl.pt
http://maxwell.rnl.ist.utl.pt

Equipa de Administração de Sistemas
Rede das Novas Licenciaturas (RNL)
Instituto Superior Técnico
http://www.rnl.ist.utl.pt
http://mega.ist.utl.pt

Attachment: pgppSZnUf854m.pgp
Description: PGP signature