Re: [PATCH tip/core/rcu 03/28] rcu: Streamline code produced by __rcu_read_unlock()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Jun 10, 2011 at 03:14:29PM +0800, Lai Jiangshan wrote:
> On 06/09/2011 03:29 AM, Paul E. McKenney wrote:
> > Given some common flag combinations, particularly -Os, gcc will inline
> > rcu_read_unlock_special() despite its being in an unlikely() clause.
> > Use noline to prohibit this misoptimization.
> > 
> > In addition, move the second barrier() in __rcu_read_unlock() so that
> > it is not on the common-case code path.  This will allow the compiler to
> > generate better code for the common-case path through __rcu_read_unlock().
> > 
> > Finally, fix up whitespace in kernel/lockdep.c to keep checkpatch happy.
> > 
> > Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
> > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > ---
> >  kernel/rcutree_plugin.h |   12 ++++++------
> >  1 files changed, 6 insertions(+), 6 deletions(-)
> > 
> > diff --git a/kernel/rcutree_plugin.h b/kernel/rcutree_plugin.h
> > index ea2e2fb..40a6db7 100644
> > --- a/kernel/rcutree_plugin.h
> > +++ b/kernel/rcutree_plugin.h
> > @@ -284,7 +284,7 @@ static struct list_head *rcu_next_node_entry(struct task_struct *t,
> >   * notify RCU core processing or task having blocked during the RCU
> >   * read-side critical section.
> >   */
> > -static void rcu_read_unlock_special(struct task_struct *t)
> > +static noinline void rcu_read_unlock_special(struct task_struct *t)
> >  {
> >  	int empty;
> >  	int empty_exp;
> > @@ -387,11 +387,11 @@ void __rcu_read_unlock(void)
> >  	struct task_struct *t = current;
> >  
> >  	barrier();  /* needed if we ever invoke rcu_read_unlock in rcutree.c */
> > -	--t->rcu_read_lock_nesting;
> > -	barrier();  /* decrement before load of ->rcu_read_unlock_special */
> > -	if (t->rcu_read_lock_nesting == 0 &&
> > -	    unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
> > -		rcu_read_unlock_special(t);
> > +	if (--t->rcu_read_lock_nesting == 0) {
> 
> > +		barrier();  /* decr before ->rcu_read_unlock_special load */
> 
> Since ACCESS_ONCE() is used for loading ->rcu_read_unlock_special, is the previous
> barrier() still needed?

It doesn't really matter until we can inline __rcu_read_unlock(), but
hopefully that day is coming soon.  So...

The concern is for cases where the compiler can see __rcu_read_lock() and
__rcu_read_unlock().  The compiler would then be within its rights to
cancel the increments and decrements of t->rcu_read_lock_nesting against
each other, which could turn a loop containing an RCU read-side critical
section into one big long critical section.

We could do --ACCESS_ONCE(t->rcu_read_lock_nesting), but that generates
lousy code on x86.  So, is there a way to make the compiler forget only
about t->rcu_read_lock_nesting rather than about all variables?

							Thanx, Paul

> > +		if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
> > +			rcu_read_unlock_special(t);
> > +	}
> >  #ifdef CONFIG_PROVE_LOCKING
> >  	WARN_ON_ONCE(ACCESS_ONCE(t->rcu_read_lock_nesting) < 0);
> >  #endif /* #ifdef CONFIG_PROVE_LOCKING */
>