Re: rcu warnings cause stack overflow

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Feb 03, 2012 at 10:33:35AM -0800, Paul E. McKenney wrote:
> On Fri, Feb 03, 2012 at 10:32:14AM +0100, Heiko Carstens wrote:
> > On Thu, Feb 02, 2012 at 11:11:16AM -0800, Paul E. McKenney wrote:
> > > On Thu, Feb 02, 2012 at 03:52:20PM +0100, Frederic Weisbecker wrote:
> > > > On Thu, Feb 02, 2012 at 01:27:42PM +0100, Heiko Carstens wrote:
> > > > > On Wed, Feb 01, 2012 at 04:14:48PM +0100, Frederic Weisbecker wrote:
> > > > > > > Removing the WARN_ON_ONCE will fix this and, if lockdep is turned on, still
> > > > > > > will find illegal uses. But it won't work for lockdep off configs...
> > > > > > > So we probably want something better than the patch below.
> > > > > > 
> > > > > > Ah ok. Hmm, but why are you using an exception to implement WARN_ON()
> > > > > > in s390? Is it to have a whole new stack for the warning path in order
> > > > > > to avoid stack overflow from the place that called the WARN_ON() ?
> > > > > 
> > > > > The reason was to reduce the code footprint of the WARN_ON() and also
> > > > > be able to print the register contents at the time the warning happened.
> > > > 
> > > > Ah ok, makes sense.
> > > 
> > > So Frederic should push his anti-recursion patch, then?
> > 
> > Yes, please.
> > 
> > Tested-by: Heiko Carstens <heiko.carstens@de.ibm.com>
> > 
> > It still generates recursive warnings because the WARNON_ONCE is inlined and
> > every different usage will generate an exception, but it didn't produce a
> > stack overflow anymore.
> > To avoid the recursive warning the patch below would help. Not sure if it's
> > worth it...
> > 
> > Subject: [PATCH] rcu: move rcu_is_cpu_idle() check warning into C file
> > 
> > From: Heiko Carstens <heiko.carstens@de.ibm.com>
> > 
> > rcu_read_lock() and rcu_read_unlock() generate a warning if a cpu is in
> > extended quiescant state. Since these functions are inlined this can cause
> > a lot of warnings if in the processing of the WARN_ON_ONCE() there is
> > another usage of e.g. rcu_read_lock(). To make sure we only get one
> > warning (and avoid possible stack overflows) uninline the check.
> > 
> > Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
> > ---
> >  include/linux/rcupdate.h |    9 +++++++--
> >  kernel/rcupdate.c        |    6 ++++++
> >  2 files changed, 13 insertions(+), 2 deletions(-)
> > 
> > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > index 81c04f4..9fe7be5 100644
> > --- a/include/linux/rcupdate.h
> > +++ b/include/linux/rcupdate.h
> > @@ -230,22 +230,27 @@ static inline void destroy_rcu_head_on_stack(struct rcu_head *head)
> > 
> >  #ifdef CONFIG_PROVE_RCU
> >  extern int rcu_is_cpu_idle(void);
> > +extern void rcu_warn_if_is_cpu_idle(void);
> >  #else /* !CONFIG_PROVE_RCU */
> >  static inline int rcu_is_cpu_idle(void)
> >  {
> >  	return 0;
> >  }
> > +
> > +static inline void rcu_warn_if_is_cpu_idle(void)
> > +{
> > +}
> >  #endif /* else !CONFIG_PROVE_RCU */
> > 
> >  static inline void rcu_lock_acquire(struct lockdep_map *map)
> >  {
> > -	WARN_ON_ONCE(rcu_is_cpu_idle());
> > +	rcu_warn_if_is_cpu_idle();
> 
> Thank you for the patch, but this WARN_ON_ONCE() has now been removed
> in favor of lockdep-RCU checks elsewhere.  This has the advantage of
> leveraging lockdep's splat-once and anti-recursion facilities.
> 
> So I believe that current -rcu covers this.  (And yes, I do need to
> push my most recent changes out.)

This still uncovers cases where we call rcu_read_lock() without matching
rcu_dereference(). Amongst this we have rcu_dereference_raw(), conditional
rcu_dereference() and may be cases where we simply have no rcu_dereference*
but we use rcu_read_lock() alone for some reason...