Re: [PATCH RFC] arm64: entry: PSTATE_I_SET is leaking on pseudo NMI mode

[Jinjie Ruan <ruanjinjie@huawei.com>]

在 2026/8/10 21:04, Vladimir Murzin 写道:
> On 8/10/26 13:44, Jinjie Ruan wrote:
>>> Agreed on the need for a clearer explanation.
>>>
>>> Perhaps local_irq_disable() isn't the right API here, and we should
>>> use local_daif_restore(DAIF_PROCCTX_NOIRQ) instead?
>> The local_irq_disable() here is pair with the local_irq_enable() in
>> preempt_schedule_irq(). So I think it is not correct to replace them
>> separately.
>>
>> The pseudo NMI does not consider how to do local_irq_disable() /enable()
>> when the interrupt is masked by the DAIF.I bit as I reported before.
> 
> Unless I missed anything local_daif_restore(DAIF_PROCCTX_NOIRQ) would
> clear DAIF.{IF} and set PMR to GIC_PRIO_IRQOFF which is the state we

Hi, Vladimir,

Yes, local_daif_restore(DAIF_PROCCTX_NOIRQ) does indeed clear the
DAIF.IF bit and then set the PMR to IRQOFF.

But there is a side effect that will clear the DAIF.{DA} bit. This will
cause `arch_irqentry_exit_need_resched()` to bypass the DAIF check.
Originally, it could not be preempted, but after this fix, it becomes
preemptible.

As the below comment said, If DAIF.{DA} is set, we may be handling an
NMI or debug context, and preemption should not occur.

30 static inline bool arch_irqentry_exit_need_resched(void)
 31 {
 32 >-------/*
 33 >------- * DAIF.DA are cleared at the start of IRQ/FIQ handling, and
when GIC
 34 >------- * priority masking is used the GIC irqchip driver will
clear DAIF.IF
 35 >------- * using gic_arch_enable_irqs() for normal IRQs. If anything
is set in
 36 >------- * DAIF we must have handled an NMI, so skip preemption.
 37 >------- */
 38 >-------if (system_uses_irq_prio_masking() && read_sysreg(daif))
 39 >------->-------return false;

Best regards,
Jinjie

> expect from local_irq_disable() so we can do local_irq_{en,dis}able()
> later on.
> 
> Cheers
> Vladimir
>