Re: [PATCH] jump_label_inc may return before the code is patched

[Gleb Natapov <gleb@redhat.com>]

On Tue, Oct 18, 2011 at 04:54:41PM -0400, Jason Baron wrote:
> On Tue, Oct 18, 2011 at 07:55:51PM +0200, Gleb Natapov wrote:
> > If cpu A calls jump_label_inc() just after atomic_add_return() is
> > called by cpu B, atomic_inc_not_zero() will return value greater then
> > zero and jump_label_inc() will return to a caller before jump_label_update()
> > finishes its job on cpu B.
> > 
> > Signed-off-by: Gleb Natapov <gleb@redhat.com>
> > diff --git a/kernel/jump_label.c b/kernel/jump_label.c
> > index a8ce450..e6f1f24 100644
> > --- a/kernel/jump_label.c
> > +++ b/kernel/jump_label.c
> > @@ -66,8 +66,9 @@ void jump_label_inc(struct jump_label_key *key)
> >  		return;
> >  
> >  	jump_label_lock();
> > -	if (atomic_add_return(1, &key->enabled) == 1)
> > +	if (atomic_read(&key->enabled) == 0)
> >  		jump_label_update(key, JUMP_LABEL_ENABLE);
> > +	atomic_inc(&key->enabled);
> >  	jump_label_unlock();
> >  }
> >  
> 
> agreed, we shouldn't return before the update happens...did this cause any
> actual problem in practice? Or just observed from code inspection?
> 
It did. I am debugging strange oopses in perf code that happen when I
create and destroy hw events on multiple cpus in parallel. At some point
I started to suspect that jump labels are not working properly and
printks confirmed that sometimes static_branch() is not taken when
->enabled is not zero. I compiled kernel without jump label support and,
as far as my testing goes, oopses disappeared. Then I found the bug by
looking at the code. Strangely enough this patch alone didn't fix oopses
yet, so some other problems are probably lurking somewhere. I am going
on vocation till next week. Will debug furtherer upon return.

> Acked-by: Jason Baron <jbaron@redhat.com>
> 
> Thanks,
> 
> -Jason

--
			Gleb.