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

[Peter Zijlstra]

On Wed, Oct 21, 2020 at 11:27:57AM -0400, Steven Rostedt wrote:
> 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?

Bah, I can't read. So I was looking at:

	if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))

but that wasn't what I actually moved around. *sigh*..

A well, I'll just remove the __ here. It's not like we super care about
performance here.

Something like so then..

---
Subject: lockdep: Fix preemption WARN for spurious IRQ-enable
From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Date: Thu Oct 22 12:23:02 CEST 2020

It is valid (albeit uncommon) to call local_irq_enable() without first
having called local_irq_disable(). In this case we enter
lockdep_hardirqs_on*() with IRQs enabled and trip a preemption warning
for using __this_cpu_read().

Use this_cpu_read() instead to avoid the warning.

Fixes: 4d004099a6 ("lockdep: Fix lockdep recursion")
Reported-by: syzbot+53f8ce8bbc07924b6417@xxxxxxxxxxxxxxxxxxxxxxxxx
Reported-by: kernel test robot <lkp@xxxxxxxxx>
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
--- a/kernel/locking/lockdep.c
+++ b/kernel/locking/lockdep.c
@@ -4057,7 +4057,7 @@ void lockdep_hardirqs_on_prepare(unsigne
 	if (unlikely(in_nmi()))
 		return;
 
-	if (unlikely(__this_cpu_read(lockdep_recursion)))
+	if (unlikely(this_cpu_read(lockdep_recursion)))
 		return;
 
 	if (unlikely(lockdep_hardirqs_enabled())) {
@@ -4126,7 +4126,7 @@ void noinstr lockdep_hardirqs_on(unsigne
 		goto skip_checks;
 	}
 
-	if (unlikely(__this_cpu_read(lockdep_recursion)))
+	if (unlikely(this_cpu_read(lockdep_recursion)))
 		return;
 
 	if (lockdep_hardirqs_enabled()) {