[RFC PATCH 30/36] arm64: irq: Report FEAT_NMI masking local IRQs

Vladimir Murzin vladimir.murzin at arm.com
Wed Jul 22 06:55:52 PDT 2026


On 7/21/26 10:06, Jinjie Ruan wrote:
> 
> On 7/9/2026 8:13 PM, Vladimir Murzin wrote:
>> From: Ada Couprie Diaz <ada.coupriediaz at arm.com>
>>
>> As we clear SCTLR_EL1.SPINTMASK when enabling FEAT_NMI, ALLINT masks
>> IRQs and FIQs regardless of superpriority.
>> Update irqflags.h and `regs_irqs_disabled()` to take it into account,
>> so we properly keep track of
>>
>> We have documentation at the top of irqflags.h which explains the DAIF
>> masking. Since the additional masking with NMIs is related and also
>> covers the IF in DAIF, extend the comment to note what's going on
>> with NMIs.
>>
>> Co-developed-by: Mark Brown <broonie at kernel.org>
>> Signed-off-by: Mark Brown <broonie at kernel.org>
>> Signed-off-by: Ada Couprie Diaz <ada.coupriediaz at arm.com>
>> Signed-off-by: Vladimir Murzin <vladimir.murzin at arm.com>
>> ---
>>  arch/arm64/include/asm/irqflags.h | 17 +++++++++++++++++
>>  arch/arm64/include/asm/ptrace.h   |  9 ++++++++-
>>  2 files changed, 25 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/arm64/include/asm/irqflags.h b/arch/arm64/include/asm/irqflags.h
>> index b571d34bf11d..a291a231d644 100644
>> --- a/arch/arm64/include/asm/irqflags.h
>> +++ b/arch/arm64/include/asm/irqflags.h
>> @@ -20,6 +20,16 @@
>>   * always masked and unmasked together, and have no side effects for other
>>   * flags. Keeping to this order makes it easier for entry.S to know which
>>   * exceptions should be unmasked.
>> + *
>> + * With the addition of the FEAT_NMI extension we gain an additional
>> + * class of superpriority IRQ/FIQ which is separately masked with a
>> + * choice of modes controlled by SCTLR_ELn.{SPINTMASK,NMI}.
>> + * Linux sets SPINTMASK to 0 and NMI to 1 which results in ALLINT.ALLINT
>> + * masking both superpriority interrupts and IRQ/FIQ regardless of the
>> + * I and F settings. Since these superpriority interrupts are being
>> + * used as NMIs we do not include them in the interrupt masking here,
>> + * anything that requires that NMIs be masked needs to explicitly do so,
>> + * but we do check for ALLINT masking IRQs/FIQs.
>>   */
>>  
>>   /*
>> @@ -103,6 +113,9 @@ static __always_inline unsigned long arch_local_save_flags(void)
>>  {
>>  	arm64_exc_hwstate_t hwstate = { .daif = read_sysreg(daif) };
>>  
>> +	if (system_uses_nmi())
>> +		hwstate.allint = read_sysreg_s(SYS_ALLINT);
>> +
>>  	if (system_uses_irq_prio_masking())
>>  		hwstate.pmr = read_sysreg_s(SYS_ICC_PMR_EL1);
>>  
>> @@ -117,6 +130,10 @@ static __always_inline bool arch_irqs_disabled_flags(unsigned long flags)
>>  	if (hwstate.daif & PSR_I_BIT)
>>  		return true;
>>  
>> +	/* SCTLR_EL1.SPINTMASK is clear, so ALLINT masks *all* IRQs/FIQs. */
>> +	if (system_uses_nmi() && hwstate.allint > 0)
> The ">0" here doesn't seem very rigorous, maybe:
> 
> if (system_uses_nmi() && hwstate.allint & ALLINT_ALLINT)
> 

Agree.

>> +		return true;
>> +
>>  	if (system_uses_irq_prio_masking() && hwstate.pmr < GIC_PRIO_IRQON)
>>  		return true;
>>  
>> diff --git a/arch/arm64/include/asm/ptrace.h b/arch/arm64/include/asm/ptrace.h
>> index c630165882e8..096d85d75bd7 100644
>> --- a/arch/arm64/include/asm/ptrace.h
>> +++ b/arch/arm64/include/asm/ptrace.h
>> @@ -206,9 +206,16 @@ static inline void forget_syscall(struct pt_regs *regs)
>>  		(regs)->pmr >= GIC_PRIO_IRQON :				\
>>  		true)
>>  
>> +#define irqs_allint_clear(regs)				\
>> +	(system_uses_nmi() ?				\
>> +		((regs)->pstate & PSR_ALLINT_BIT) == 0:	\
>> +		true)
>> +
>>  static __always_inline bool regs_irqs_disabled(const struct pt_regs *regs)
>>  {
>> -	return (regs->pstate & PSR_I_BIT) || !irqs_priority_unmasked(regs);
>> +	return (regs->pstate & PSR_I_BIT) ||
>> +		!irqs_priority_unmasked(regs) ||
>> +		!irqs_allint_clear(regs);
> Why not define the logic for disable? The semantics of
> regs_irqs_disabled() is "interrupts disabled".
> 
> 204 #define
> irqs_priority_masked(regs)>----->------->------->------->-------\
> 205 >-------(system_uses_irq_prio_masking()
> ?>------>------->------->-------\
> 206 >------->-------(regs)->pmr == GIC_PRIO_IRQOFF
> :>------->------->-------\
> 207 >------->-------false)
> 208
> 209 #define irqs_allint_masked(regs)>------->------->------->-------\
> 210 >-------(system_uses_nmi() ?>--->------->------->-------\
> 211 >------->-------((regs)->pstate & PSR_ALLINT_BIT) :>----\
> 212 >------->-------false)
> 213
> 214 static __always_inline bool regs_irqs_disabled(const struct pt_regs
> *regs)
> 215 {
> 216 >-------return (regs->pstate & PSR_I_BIT) ||
> 217 >------->-------irqs_priority_masked(regs) ||
> 218 >------->-------irqs_allint_masked(regs);
> 

That's in line with my proposal [1], with that it would look like

static __always_inline bool regs_irqs_disabled(const struct pt_regs *regs)
{
        if (regs->pstate & PSR_I_BIT)
                return true;

        if (system_uses_nmi())
                return !!(regs->pstate & PSR_ALLINT_BIT);

        if (system_uses_irq_prio_masking()) {
                WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_IRQFLAGS) &&
                             regs->pmr != GIC_PRIO_IRQOFF &&
                             regs->pmr != GIC_PRIO_IRQON);

                return regs->pmr != GIC_PRIO_IRQON;
        }

        return false;
}


[1] https://lore.kernel.org/linux-arm-kernel/fedf2b3e-14ed-4483-8cdf-d31f1ac97b84@arm.com/

> otherwise, LGTM
> Reviewed-by: Jinjie Ruan <ruanjinjie at huawei.com>
> 
>>  }
>>  
>>  #define interrupts_enabled(regs)	(!regs_irqs_disabled(regs))
> 

Cheers
Vladimir



More information about the linux-arm-kernel mailing list