Re: [PATCH v7 11/25] arm64: irqflags: Use ICC_PMR_EL1 for interrupt masking

[Jian-Lin Chen]

HI Julien,

Thanks a lot for your reply, since I'm working on this patch in ARM
(32 bits), so I have to dig into the details.

Julien Thierry <julien.thierry@xxxxxxx> æ 2018å12æ17æ éä äå5:26åéï
>
> Hi Jian-Lin,
>
> Thanks for looking at this.
>
> On 16/12/2018 14:47, Jian-Lin Chen wrote:
> > From: Jian-Lin Chen <lecopzer.chen@xxxxxxxxxxxx>
> >
> >
> > On Wed, 12 Dec 2018 at 17:48, Julien Thierry <julien.thierry@xxxxxxx> wrote:
> >>  static inline void arch_local_irq_enable(void)
> >>  {
> >> -       asm volatile(
> >> -               "msr    daifclr, #2             // arch_local_irq_enable"
> >> -               :
> >> +       unsigned long unmasked = GIC_PRIO_IRQON;
> >> +
> >
> > Should we need a WARN_ON() to check if the daif_I bit is masked, or
> > explicitly unmasked I bit here?
> >
>
> While I would agree, adding the WARN_ON() will add some non-negligible
> overhead, especially if we need to read the daif flags to check it.
>
> Since these functions are called often in the whole system and using PMR
> already makes things a bit slower, I'd prefer to avoid checks in here.

Ok, so we have to find a better place to check it.
I have no idea so far...


>
> > If I bit was masked and someone calls arch_local_irq_enable(), they still
> > couldn't recieve any interrupt.
> >
> >
> >> +       asm volatile(ALTERNATIVE(
> >> +               "msr    daifclr, #2             // arch_local_irq_enable\n"
> >> +               "nop",
> >> +               "msr_s  " __stringify(SYS_ICC_PMR_EL1) ",%0\n"
> >> +               "dsb    sy",
> >> +               ARM64_HAS_IRQ_PRIO_MASKING)
> >>                 :
> >> +               : "r" (unmasked)
> >>                 : "memory");
> >>  }
> >>
> >>  static inline void arch_local_irq_disable(void)
> >>  {
> >> -       asm volatile(
> >> -               "msr    daifset, #2             // arch_local_irq_disable"
> >> -               :
> >> +       unsigned long masked = GIC_PRIO_IRQOFF;
> >> +
> >> +       asm volatile(ALTERNATIVE(
> >> +               "msr    daifset, #2             // arch_local_irq_disable",
> >> +               "msr_s  " __stringify(SYS_ICC_PMR_EL1) ", %0",
> >
> > May be a "dsb sy" here?
>
> So, we need a "dsb sy" when unmasking interrupts because this ensures
> the redistributor sees the latest PMR value and starts forwarding lower
> priority interrupts again.
>
> When we disable interrupts however, the GIC CPU interface guarantees
> that no interrupts of lower priority than the current value of PMR will
> be taken. So we don't really need the redistributor to immediately see
> the new value of PMR as the logic in the GIC CPU interface is good
> enough for our goal.
>

Got it, thanks for the detail!



> Thanks,
>
> --
> Julien Thierry