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

[peterz at infradead.org]

On Wed, Aug 19, 2020 at 05:32:50PM +0200, peterz at infradead.org wrote:
> On Wed, Aug 19, 2020 at 08:39:13PM +1000, Alexey Kardashevskiy wrote:
> 
> > > or current upstream?
> > 
> > The upstream 18445bf405cb (13 hours old) also shows the problem. Yours
> > 1/2 still fixes it.
> 
> Afaict that just reduces the window.
> 
> Isn't the problem that:
> 
> arch/powerpc/kernel/exceptions-64e.S
> 
> 	START_EXCEPTION(perfmon);
> 	NORMAL_EXCEPTION_PROLOG(0x260, BOOKE_INTERRUPT_PERFORMANCE_MONITOR,
> 				PROLOG_ADDITION_NONE)
> 	EXCEPTION_COMMON(0x260)
> 	INTS_DISABLE
> #	  RECONCILE_IRQ_STATE
> #	    TRACE_DISABLE_INTS
> #	      TRACE_WITH_FRAME_BUFFER(trace_hardirqs_off)
> #
> # but we haven't done nmi_enter() yet... whoopsy
> 
> 	CHECK_NAPPING()
> 	addi	r3,r1,STACK_FRAME_OVERHEAD
> 	bl	performance_monitor_exception
> #	 perf_irq()
> #          perf_event_interrupt
> #	     __perf_event_interrupt
> #	      nmi_enter()
> 
> 
> 
> That is, afaict your entry code is buggered.

That is, patch 1/2 doesn't change the case:

	local_irq_enable()
	  trace_hardirqs_on()
	  <NMI>
	    trace_hardirqs_off()
	    ...
	    if (regs_irqs_disabled(regs)) // false
	      trace_hardirqs_on();
	  </NMI>
	  raw_local_irq_enable()

Where local_irq_enable() has done trace_hardirqs_on() and the NMI hits
and undoes it, but doesn't re-do it because the hardware state is still
disabled.

What's supposed to happen is:

	<NMI>
	  nmi_enter()
	  trace_hardirqs_off() // no-op, because in_nmi() (or previously because lockdep_off())
	  ...
	</NMI>