Re: [BUG] 2.6.37-rc3 massive interactivity regression on ARM

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2010-12-09 at 10:11 -0800, Venkatesh Pallipadi wrote:
> On Thu, Dec 9, 2010 at 9:55 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > On Thu, 2010-12-09 at 09:43 -0800, Venkatesh Pallipadi wrote:
> >>
> >> The same problem will be there with below code, with irq_delta >
> >> delta, clock_task can go backwards which is not good.
> >> +       delta -= irq_delta;
> >> +       rq->clock_task += delta;
> >>
> >> The reason for this is rq->clock and irqtime updates kind of happen
> >> independently and specifically, if a rq->clock update happens while we
> >> are in a softirq, we may have this case of going backwards on the next
> >> update.
> >
> > But how can irq_delta > delta?, we measure it using the same clock.
> >
> 
> This would be mostly a corner case like:
> - softirq start time t1
> - rq->clock updated at t2 and rq->clock_task updated at t2 without
> accounting for current softirq
> - softirq end time t3
> - cpu spends most time here in softirq or hardirq
> - next rq->clock update at t4 and rq->clock_task update, with delta =
> t4-t2 and irq_delta ~= t4 - t1

Ah, something like that would happen when we do a wakeup from
soft/hard-irq context, not an altogether uncommon occurrence.

Wouldn't that be cured by updating the irq-time when asking for it,
something like the below? (on top of my earlier patch)

---
Index: linux-2.6/kernel/sched.c
===================================================================
--- linux-2.6.orig/kernel/sched.c
+++ linux-2.6/kernel/sched.c
@@ -1843,8 +1843,9 @@ void account_system_vtime(struct task_st
 }
 EXPORT_SYMBOL_GPL(account_system_vtime);
 
-static inline u64 irq_time_cpu(int cpu)
+static u64 irq_time_cpu(int cpu)
 {
+	account_system_vtime(current);
 	return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
 }