Re: BUG: using __this_cpu_read() in preemptible code in trace_hardirqs_on

[Steven Rostedt]

On Wed, 21 Oct 2020 17:12:37 +0200
Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> > > diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
> > > index 3e99dfef8408..9f818145ef7d 100644
> > > --- a/kernel/locking/lockdep.c
> > > +++ b/kernel/locking/lockdep.c
> > > @@ -4057,9 +4057,6 @@ void lockdep_hardirqs_on_prepare(unsigned long ip)
> > >  	if (unlikely(in_nmi()))
> > >  		return;
> > >  
> > > -	if (unlikely(__this_cpu_read(lockdep_recursion)))
> > > -		return;
> > > -
> > >  	if (unlikely(lockdep_hardirqs_enabled())) {  
> > 
> > Hmm, would moving the recursion check below the check of the
> > lockdep_hardirqs_enable() cause a large skew in the spurious enable stats?
> > May not be an issue, but something we should check to make sure that
> > there's not a path that constantly hits this.  
> 
> Anything that sets recursion will have interrupts disabled.

It may have interrupts disabled, but does it have the hardirqs_enabled
per_cpu variable set? The above check only looks at that, and doesn't check
if interrupts are actually enabled.

For example, if lockdep is processing a mutex, it would set the recursion
variable, but does it ever set the hardirqs_enabled variable to off?

-- Steve