Re: [tip:sched/urgent] sched: Check sched_domain before computing group power

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Nov 21, 2013 at 09:22:24AM -0800, Yinghai Lu wrote:
> On Thu, Nov 21, 2013 at 7:03 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> >>
> >> This one seems fix NULL reference in compute_group_power.
> >>
> >> but get following on current Linus tree plus tip/sched/urgent.
> >>
> >> divide error: 0000 [#1]  SMP
> >> [   28.190477] Modules linked in:
> >> [   28.192012] CPU: 11 PID: 484 Comm: kworker/u324:0 Not tainted
> >> 3.12.0-yh-10487-g4b94e59-dirty #2044
> >> [   28.210488] Hardware name: Oracle Corporation  Sun Fire
> >> [   28.229877] task: ffff88ff25205140 ti: ffff88ff2520a000 task.ti:
> >> ffff88ff2520a000
> >> [   28.236139] RIP: 0010:[<ffffffff810d9ff4>]  [<ffffffff810d9ff4>]
> >> find_busiest_group+0x2b4/0x8a0
> >
> > Hurmph.. what kind of hardware is that? and is there anything funny you
> > do to make it do this?
> 
> intel nehanem-ex or westmere-ex 8 sockets system.
> 
> I tried without my local patches, the problem is still there.

And I suppose a kernel before

  863bffc80898 ("sched/fair: Fix group power_orig computation")

work fine, eh?

I'll further assume that your RIP points to:

	sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;

indicating that sds.total_pwr := 0.

update_sd_lb_stats() computes it like:

		sds->total_pwr += sgs->group_power;

which comes out of update_sg_lb_stats() like:

	sgs->group_power = group->sgp->power;

Which we compute in update_group_power() similarly to how we did before
863bffc80898.

Which leaves me a bit puzzled.