On 05/12/18 18:26, Catalin Marinas wrote:
> On Wed, Dec 05, 2018 at 04:55:54PM +0000, Julien Thierry wrote:
>> On 04/12/18 17:36, Catalin Marinas wrote:
>>> On Mon, Nov 12, 2018 at 11:57:01AM +0000, Julien Thierry wrote:
>>>> diff --git a/arch/arm64/include/asm/irqflags.h
>>>> b/arch/arm64/include/asm/irqflags.h
>>>> index 24692ed..e0a32e4 100644
>>>> --- a/arch/arm64/include/asm/irqflags.h
>>>> +++ b/arch/arm64/include/asm/irqflags.h
>>>> @@ -18,7 +18,27 @@
>>>>
>>>> #ifdef __KERNEL__
>>>>
>>>> +#include <asm/alternative.h>
>>>> +#include <asm/cpufeature.h>
>>>> #include <asm/ptrace.h>
>>>> +#include <asm/sysreg.h>
>>>> +
>>>> +
>>>> +/*
>>>> + * When ICC_PMR_EL1 is used for interrupt masking, only the bit indicating
>>>> + * whether the normal interrupts are masked is kept along with the daif
>>>> + * flags.
>>>> + */
>>>> +#define ARCH_FLAG_PMR_EN 0x1
>>>> +
>>>> +#define MAKE_ARCH_FLAGS(daif, pmr)
>>>> \
>>>> + ((daif) | (((pmr) >> GIC_PRIO_STATUS_SHIFT) & ARCH_FLAG_PMR_EN))
>>>> +
>>>> +#define ARCH_FLAGS_GET_PMR(flags) \
>>>> + ((((flags) & ARCH_FLAG_PMR_EN) << GIC_PRIO_STATUS_SHIFT) \
>>>> + | GIC_PRIO_IRQOFF)
>>>> +
>>>> +#define ARCH_FLAGS_GET_DAIF(flags) ((flags) & ~ARCH_FLAG_PMR_EN)
>>>
>>> I wonder whether we could just use the PSR_I_BIT here to decide whether
>>> to set the GIC_PRIO_IRQ{ON,OFF}. We could clear the PSR_I_BIT in
>>> _restore_daif() with an alternative.
>>
>> So, the issue with it is that some contexts might be using PSR.I to
>> disable interrupts (any contexts with async errors or debug exceptions
>> disabled, kvm guest entry paths, pseudo-NMIs, ...).
>>
>> If any of these contexts calls local_irq_save()/local_irq_restore() or
>> local_daif_save()/local_daif_restore(), by only relying on PSR_I_BIT to
>> represent the PMR status, we might end up clearing PSR.I when we shouldn't.
>>
>> I'm not sure whether there are no callers of these functions in those
>> context. But if that is the case, we could simplify things, yes.
>
> There are callers of local_daif_save() (3) and local_daif_mask() (7) but
> do they all need to disable the pseudo-NMIs?
>
Hmmm, I really think that both of those should be disabling NMIs.
Otherwise, if we take an NMI, the first thing the el1_irq handler is
going to do is "enable_da_f()" which could lead to potential issues.
One thing that could be done is:
- local_daif_save() and local_daif_mask() both mask all daif bits
(taking care to represent PMR value in the I bit of the saved flags)
- local_daif_restore() restores da_f as expected and decides values to
put for PMR and PSR.I as follows:
* do the da_f restore
* if PSR.A bit is cleared in the saved flags, then we also do a
start_nmi()
However, this would not work with a local_daif_save()/restore() on the
return path of an NMI because I think it is the only context with NMIs
"stopped" that can take aborts. I can add a WARN_ON(in_nmi()) for
local_daif_restore() if that doesn't affect performance too much.
Does that sound alright?
> At a brief look at x86, it seems that they have something like
> stop_nmi() and restart_nmi(). These don't have save/restore semantics,
> so we could do something similar on arm64 that only deals with the
> PSTATE.I bit directly and keep the software (flags) PSR.I as the PMR
> bit. But we'd have to go through the 10 local_daif_* cases above to see
> which actually need the stop_nmi() semantics.
>
Yes, having those could be useful to deal with the above and maybe some
other places.
Thanks,
--
Julien Thierry