[BUG] 2.6.37-rc3 massive interactivity regression on ARM

[Venkatesh Pallipadi]

On Thu, Dec 9, 2010 at 3:16 PM, Peter Zijlstra <peterz at infradead.org> wrote:
> On Thu, 2010-12-09 at 14:21 -0800, Venkatesh Pallipadi wrote:
>> This should mostly work. There would be a small window there on
>> rq->clock=sched_clock_cpu() and system_vtime doing sched_clock_cpu()
>> and also the overhead of doing this everytime when this going
>> backwards may happen rarely.
>
> Right, so something like the below removes that tiny race by re-using
> rq->clock as now. It also removes some overhead by removing the IRQ
> fudging (we already know IRQs are disabled).
>
> It does have a few extra branches (2 afaict) than your monotonicity path
> had, but it seems to me this approach is slightly more accurate.
>

Looks fine.

Just to make sure, update_rq_clock() always gets called on current
CPU. Right? The pending patches I have optimizes
account_system_vtime() to use this_cpu_write and friends. Want to make
sure this change will still keep that optimization relevant.


> ---
> Index: linux-2.6/kernel/sched.c
> ===================================================================
> --- linux-2.6.orig/kernel/sched.c
> +++ linux-2.6/kernel/sched.c
> @@ -1813,19 +1813,10 @@ void disable_sched_clock_irqtime(void)
>        sched_clock_irqtime = 0;
>  }
>
> -void account_system_vtime(struct task_struct *curr)
> +static void __account_system_vtime(int cpu, u64 now)
>  {
> -       unsigned long flags;
> -       int cpu;
> -       u64 now, delta;
> +       s64 delta;
>
> -       if (!sched_clock_irqtime)
> -               return;
> -
> -       local_irq_save(flags);
> -
> -       cpu = smp_processor_id();
> -       now = sched_clock_cpu(cpu);
>        delta = now - per_cpu(irq_start_time, cpu);
>        per_cpu(irq_start_time, cpu) = now;
>        /*
> @@ -1836,16 +1827,36 @@ void account_system_vtime(struct task_st
>         */
>        if (hardirq_count())
>                per_cpu(cpu_hardirq_time, cpu) += delta;
> -       else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
> +       else if (in_serving_softirq() && !(current->flags & PF_KSOFTIRQD))
>                per_cpu(cpu_softirq_time, cpu) += delta;
> +}
> +
> +void account_system_vtime(struct task_struct *curr)
> +{
> +       unsigned long flags;
> +       u64 now;
> +       int cpu;
> +
> +       if (!sched_clock_irqtime)
> +               return;
> +
> +       local_irq_save(flags);
> +
> +       cpu = smp_processor_id();
> +       now = sched_clock_cpu(cpu);
> +       __account_system_vtime(cpu, now);
>
>        local_irq_restore(flags);
>  }
>  EXPORT_SYMBOL_GPL(account_system_vtime);
>
> -static u64 irq_time_cpu(int cpu)
> +static u64 irq_time_cpu(struct rq *rq)
>  {
> -       account_system_vtime(current);
> +       int cpu = cpu_of(rq);
> +
> +       if (sched_clock_irqtime)
> +               __account_system_vtime(cpu, rq->clock);
> +
>        return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
>  }
>
> @@ -1853,7 +1864,7 @@ static void update_rq_clock_task(struct
>  {
>        s64 irq_delta;
>
> -       irq_delta = irq_time_cpu(cpu_of(rq)) - rq->prev_irq_time;
> +       irq_delta = irq_time_cpu(rq) - rq->prev_irq_time;
>        rq->prev_irq_time += irq_delta;
>
>        delta -= irq_delta;
>
>