Re: rcu-bh design

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

On Fri, May 04, 2018 at 08:33:19PM +0000, Joel Fernandes wrote:
> On Fri, May 4, 2018 at 1:10 PM Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> wrote:
> [...]
> > > >> > Almost.  All context switches in an RCU-preempt read-side critical
> section
> > > >> > must be subject to priority boosting.  Preemption is one example,
> because
> > > >> > boosting the priority of the preempted task will make it runnable.
> > > >> > The priority-inheritance -rt "spinlock" is another example, because
> > > >> > boosting the priority of the task holding the lock will eventually
> make
> > > >> > runnable the task acquiring the lock within the RCU-preempt
> read-side
> > > >> > critical section.
> > > >>
> > > >> Yes I understand priority boosting is needed with preemptible RCU so
> that
> > > >> read-sections are making forward progress. I meant (and correct me
> if I'm
> > > >> wrong) that, as long as a task doesn't sleep in a preemptible RCU
> > > >> read-section (rcu-preempt flavor), then bad things wont happen and
> RCU will
> > > >> work correctly.
> > > >
> > > > The exception is -rt "spinlock" acquisition.  If the "spinlock" is
> held,
> > > > the task acquiring it will block, which is legal within an RCU-preempt
> > > > read-side critical section.
> > > >
> > > > This exception is why I define bad things in terms of lack of
> > > > susceptibility to priority boosting instead of sleeping.
> > >
> > > Oh, that's a tricky situation. Thanks for letting me know. I guess my
> > > view was too idealistic. Makes sense now.
> 
> > Well, let me put it this way...
> 
> > Here is your nice elegant little algorithm:
> > https://upload.wikimedia.org/wikipedia/commons/6/6e/Golde33443.jpg
> 
> > Here is your nice elegant little algorithm equipped to survive within
> > the Linux kernel:
> > https://totatema.files.wordpress.com/2012/06/feeling_grizzly-1600x1200.jpg
> 
> A picture speaks a 1000 words! :-D

And I suspect that a real face-to-face encounter with that guy is worth
1000 pictures.  ;-)

> > Any questions?  ;-)
> 
> Yes just one more ;-). I am trying to write a 'probetorture' test inspired
> by RCU torture that whacks the tracepoints in various scenarios. One of the
> things I want to do is verify the RCU callbacks are queued and secondly,
> they are executed. Just to verify that the garbage collect was done and
> we're not leaking the function probe table (not that I don't have
> confidence in the chained callback technique which you mentioned, but it
> would be nice to assure this mechanism is working for tracepoints).
> 
> Is there a way to detect this given a reference to srcu_struct? Mathieu and
> we were chatting about srcu_barrier which is cool but that just tells me
> that if there was a callback queued, it would have executed after the
> readers were done. But not whether something was queued.

Suppose that you are queuing an RCU callback that in turn queues an SRCU
callback on my_srcu_struct, like this:

	void my_rcu_callback(struct rcu_head *rhp)
	{
		p = container_of(rhp, struct my_struct, my_rcu_head);

		free_it_up_or_down(p);
	}

	void my_srcu_callback(struct rcu_head *rhp)
	{
		call_rcu(rhp, my_rcu_callback);
	}

	call_srcu(&my_srcu_struct, &p->my_rcu_head, my_srcu_callback);

Then to make sure that any previously submitted callback has been fully
processed, you do this:

	rcu_barrier();
	srcu_barrier(&my_srcu_struct);

Of course if you queue in the opposite order, like this:

	void my_srcu_callback(struct rcu_head *rhp)
	{
		p = container_of(rhp, struct my_struct, my_rcu_head);

		free_it_up_or_down(p);
	}

	void my_rcu_callback(struct rcu_head *rhp)
	{
		call_srcu(&my_srcu_struct, &p->my_rcu_head, my_srcu_callback);
	}

	call_rcu(rhp, my_rcu_callback);

Then you must wait in the opposite order:

	rcu_barrier();
	srcu_barrier(&my_srcu_struct);

Either way, the trick is that the first *_barrier() call cannot return
until after all previous callbacks have executed, which means that by that
time the callback is enqueued for the other flavor of {S,}RCU.  So the
second *_barrier() call must wait for the callback to be completely done,
through both flavors of {S,}RCU.

So after executing the pair of *_barrier() calls, you know that the
callback is no longer queued.

Does that cover it, or am I missing a turn in here somewhere?

							Thanx, Paul

> thanks,
> 
> - Joel
> PS: I remember Paul, you mentioned you are testing this chained callback
> case in rcutorture, so if the answer is "go read rcutorture", that's
> totally Ok I could just do that ;-)
>