Re: [PATCH 1/2] lockdep: improve current->(hard|soft)irqs_enabled synchronisation with actual irq state

[Peter Zijlstra]

On Thu, Jul 23, 2020 at 08:56:14PM +1000, Nicholas Piggin wrote:
> diff --git a/arch/powerpc/include/asm/hw_irq.h b/arch/powerpc/include/asm/hw_irq.h
> index 3a0db7b0b46e..35060be09073 100644
> --- a/arch/powerpc/include/asm/hw_irq.h
> +++ b/arch/powerpc/include/asm/hw_irq.h
> @@ -200,17 +200,14 @@ static inline bool arch_irqs_disabled(void)
>  #define powerpc_local_irq_pmu_save(flags)			\
>  	 do {							\
>  		raw_local_irq_pmu_save(flags);			\
> -		trace_hardirqs_off();				\
> +		if (!raw_irqs_disabled_flags(flags))		\
> +			trace_hardirqs_off();			\
>  	} while(0)

So one problem with the above is something like this:

	raw_local_irq_save();
	<NMI>
	  powerpc_local_irq_pmu_save();

that would now no longer call into tracing/lockdep at all. As a
consequence, lockdep and tracing would show the NMI ran with IRQs
enabled, which is exceptionally weird..

Similar problem with:

	raw_local_irq_disable();
	local_irq_save()

Now, most architectures today seem to do what x86 also did:

	<NMI>
	  trace_hardirqs_off()
	  ...
	  if (irqs_unmasked(regs))
	    trace_hardirqs_on()
	</NMI>

Which is 'funny' when it interleaves like:

	local_irq_disable();
	...
	local_irq_enable()
	  trace_hardirqs_on();
	  <NMI/>
	  raw_local_irq_enable();

Because then it will undo the trace_hardirqs_on() we just did. With the
result that both tracing and lockdep will see a hardirqs-disable without
a matching enable, while the hardware state is enabled.

Which is exactly the state Alexey seems to have ran into.

Now, x86, and at least arm64 call nmi_enter() before
trace_hardirqs_off(), but AFAICT Power never did that, and that's part
of the problem. nmi_enter() does lockdep_off() and that _used_ to also
kill IRQ tracking.

Now, my patch changed that, it makes IRQ tracking not respect
lockdep_off(). And that exposed x86 (and everybody else :/) to the same
problem you have.

And this is why I made x86 look at software state in NMIs. Because then
it all works out. For bonus points, trace_hardirqs_*() also has some
do-it-once logic for tracing.



Anyway, it's Saturday evening, time for a beer. I'll stare at this more
later.