Re: [PATCH tip/core/rcu 02/18] rcu: Move rcu_report_exp_rnp() to allow consolidation

[Paul E. McKenney]

On Wed, Oct 07, 2015 at 09:51:14AM +0200, Peter Zijlstra wrote:
> On Tue, Oct 06, 2015 at 01:58:50PM -0700, Paul E. McKenney wrote:
> > On Tue, Oct 06, 2015 at 10:29:37PM +0200, Peter Zijlstra wrote:
> > > On Tue, Oct 06, 2015 at 09:29:21AM -0700, Paul E. McKenney wrote:
> > > > +static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp,
> > > > +					      struct rcu_node *rnp, bool wake)
> > > > +{
> > > > +	unsigned long flags;
> > > > +	unsigned long mask;
> > > > +
> > > > +	raw_spin_lock_irqsave(&rnp->lock, flags);
> > > 
> > > Normally we require a comment with barriers, explaining the order and
> > > the pairing etc.. :-)
> > > 
> > > > +	smp_mb__after_unlock_lock();
> > 
> > Hmmmm...  That is not good.
> > 
> > Worse yet, I am missing comments on most of the pre-existing barriers
> > of this form.
> 
> Yes I noticed.. :/

Will fix, though probably as a follow-up patch.  Once I figure out what
comment makes sense...

> > The purpose is to enforce the heavy-weight grace-period memory-ordering
> > guarantees documented in the synchronize_sched() header comment and
> > elsewhere.
> 
> > They pair with anything you might use to check for violation
> > of these guarantees, or, simiarly, any ordering that you might use when
> > relying on these guarantees.
> 
> I'm sure you know what that means, but I've no clue ;-) That is, I
> wouldn't know where to start looking in the RCU implementation to verify
> the barrier is either needed or sufficient. Unless you mean _everywhere_
> :-)

Pretty much everywhere.

Let's take the usual RCU removal pattern as an example:

	void f1(struct foo *p)
	{
		list_del_rcu(p);
		synchronize_rcu_expedited();
		kfree(p);
	}

	void f2(void)
	{
		struct foo *p;

		list_for_each_entry_rcu(p, &my_head, next)
			do_something_with(p);
	}

So the synchronize_rcu_expedited() acts as an extremely heavyweight
memory barrier that pairs with the rcu_dereference() inside of
list_for_each_entry_rcu().  Easy enough, right?

But what exactly within synchronize_rcu_expedited() provides the
ordering?  The answer is a web of lock-based critical sections and
explicit memory barriers, with the one you called out as needing
a comment being one of them.

> > I could add something like  "/* Enforce GP memory ordering. */"
> > 
> > Or perhaps "/* See synchronize_sched() header. */"
> > 
> > I do not propose reproducing the synchronize_sched() header on each
> > of these.  That would be verbose, even for me!  ;-)
> > 
> > Other thoughts?
> 
> Well, this is an UNLOCK+LOCK on non-matching lock variables upgrade to
> full barrier thing, right?

Yep!

> To me its not clear which UNLOCK we even match here. I've just read the
> sync_sched() header, but that doesn't help me either, so referring to
> that isn't really helpful either.

Usually this pairs with an rcu_dereference() somewhere in the calling
code.  Some other task in the calling code, actually.

> In any case, I don't want to make too big a fuzz here, but I just
> stumbled over a lot of unannotated barriers and figured I ought to say
> something about it.

I do need to better document how this works, no two ways about it.

							Thanx, Paul

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/