On Thu, Jul 29, 2004 at 07:57:55AM -0300, Marcelo Tosatti wrote:
> On Thu, Jul 29, 2004 at 09:54:29AM +0200, Arjan van de Ven wrote:
> > On Wed, Jul 28, 2004 at 11:27:41PM -0700, Chris Caputo wrote:
> > > On Wed, 28 Jul 2004, Marcelo Tosatti wrote:
> > > > Changing the affinity writes new values to the IOAPIC registers, I can't see
> > > > how that could interfere with the atomicity of a spinlock operation. I dont
> > > > understand why you think irqbalance could affect anything.
> > > 
> > > Because when I stop running irqbalance the crashes no longer happen.
> > 
> > what is the irq distribution when you do that?
> > Can you run irqbalance for a bit to make sure there's a static distribution
> > of irq's and then disable it and see if it survives ?
> 
> Chris, Yes I'm also running irqbalance. 
> 
> Arjan, what is an easy way for me to make irqbalance change the affinity
> as crazy on the SMP 8way box, just for a test?

there is a sleep(10 seconds) in the code, if you change that to something
really short and then cause irq burst rates on different devices...