On Fri, 2018-04-06 at 00:31 +1000, Nicholas Piggin wrote:
> irq_work_raise should not schedule the hardware decrementer interrupt
> unless it is called from NMI context. Doing so often just results in an
> immediate masked decrementer interrupt:
> 
>    <...>-550    90d...    4us : update_curr_rt <-dequeue_task_rt
>    <...>-550    90d...    5us : dbs_update_util_handler <-update_curr_rt
>    <...>-550    90d...    6us : arch_irq_work_raise <-irq_work_queue
>    <...>-550    90d...    7us : soft_nmi_interrupt <-soft_nmi_common
>    <...>-550    90d...    7us : printk_nmi_enter <-soft_nmi_interrupt
>    <...>-550    90d.Z.    8us : rcu_nmi_enter <-soft_nmi_interrupt
>    <...>-550    90d.Z.    9us : rcu_nmi_exit <-soft_nmi_interrupt
>    <...>-550    90d...    9us : printk_nmi_exit <-soft_nmi_interrupt
>    <...>-550    90d...   10us : cpuacct_charge <-update_curr_rt
> 
> Set the decrementer pending in the irq_happened mask directly, rather
> than having the masked decrementer handler do it.

Setting the paca field needs hard irqs off... also preempt_disable
doesn't look necessary if IRQs are off.

> Signed-off-by: Nicholas Piggin <npig...@gmail.com>
> ---
>  arch/powerpc/kernel/time.c | 35 +++++++++++++++++++++++++++++++++--
>  1 file changed, 33 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/powerpc/kernel/time.c b/arch/powerpc/kernel/time.c
> index a32823dcd9a4..9d1cc183c974 100644
> --- a/arch/powerpc/kernel/time.c
> +++ b/arch/powerpc/kernel/time.c
> @@ -510,6 +510,35 @@ static inline void clear_irq_work_pending(void)
>               "i" (offsetof(struct paca_struct, irq_work_pending)));
>  }
>  
> +void arch_irq_work_raise(void)
> +{
> +     WARN_ON(!irqs_disabled());
> +
> +     preempt_disable();
> +     set_irq_work_pending_flag();
> +     /*
> +      * Regular iterrupts will check pending irq_happened as they return,
> +      * or process context when it next enables interrupts, so the
> +      * decrementer can be scheduled there.
> +      *
> +      * NMI interrupts do not, so setting the decrementer hardware
> +      * interrupt to fire ensures the work runs upon RI (if it's to a
> +      * MSR[EE]=1 context). We do not want to do this in other contexts
> +      * because if interrupts are hard enabled, the decrementer will
> +      * fire immediately here and just go to the masked handler to be
> +      * recorded in irq_happened.
> +      *
> +      * BookE does not support this yet, it must audit all NMI
> +      * interrupt handlers call nmi_enter().
> +      */
> +     if (IS_ENABLED(CONFIG_BOOKE) || in_nmi()) {
> +             set_dec(1);
> +     } else {
> +             local_paca->irq_happened |= PACA_IRQ_DEC;
> +     }
> +     preempt_enable();
> +}
> +
>  #else /* 32-bit */
>  
>  DEFINE_PER_CPU(u8, irq_work_pending);
> @@ -518,16 +547,18 @@ DEFINE_PER_CPU(u8, irq_work_pending);
>  #define test_irq_work_pending()              
> __this_cpu_read(irq_work_pending)
>  #define clear_irq_work_pending()     __this_cpu_write(irq_work_pending, 0)
>  
> -#endif /* 32 vs 64 bit */
> -
>  void arch_irq_work_raise(void)
>  {
> +     WARN_ON(!irqs_disabled());
> +
>       preempt_disable();
>       set_irq_work_pending_flag();
>       set_dec(1);
>       preempt_enable();
>  }
>  
> +#endif /* 32 vs 64 bit */
> +
>  #else  /* CONFIG_IRQ_WORK */
>  
>  #define test_irq_work_pending()      0

Reply via email to