Re: [PATCH 08/10] rcu: Allow to deactivate nocb on a CPU

[Frederic Weisbecker <frederic@kernel.org>]

On Thu, May 14, 2020 at 08:47:07AM -0700, Paul E. McKenney wrote:
> On Thu, May 14, 2020 at 12:45:26AM +0200, Frederic Weisbecker wrote:
> This last seems best to me.  The transition from CBLIST_NOT_OFFLOADED
> to CBLIST_OFFLOADING of course needs to be on the CPU in question with
> at least bh disabled.  Probably best to be holding rcu_nocb_lock(),
> but that might just be me being overly paranoid.

So that's in the case of offloading, right? Well, I don't think we'd
need to even disable bh nor lock nocb. We just need the current CPU
to see the local update of cblist->offloaded = CBLIST_OFFLOADING
before the kthread is unparked:

    cblist->offloaded = CBLIST_OFFLOADING;
    /* Make sure subsequent softirq lock nocb */
    barrier();
    kthread_unpark(rdp->nocb_cb_thread);

Now, although that guarantees that nocb_cb will see CBLIST_OFFLOADING
upon unparking, it's not guaranteed that the nocb_gp will see it on its
next round. Ok so eventually you're right, I should indeed lock nocb...

> 
> > > > +static long rcu_nocb_rdp_deoffload(void *arg)
> > > > +{
> > > > +	struct rcu_data *rdp = arg;
> > > > +
> > > > +	WARN_ON_ONCE(rdp->cpu != raw_smp_processor_id());
> > > > +	__rcu_nocb_rdp_deoffload(rdp);
> > > > +
> > > > +	return 0;
> > > > +}
> > > 
> > > For example, is the problem caused by invocations of this
> > > rcu_nocb_rdp_deoffload() function?
> > 
> > How so?
> 
> It looked to me like it wasn't excluding either rcu_barrier() or CPU
> hotplug.  It might also not have been pinning onto the CPU in question,
> but that might just be me misremembering.  Then again, I didn't see a
> call to it, so maybe its callers set things up appropriately.
> 
> OK, I will bite...  What is the purpose of rcu_nocb_rdp_deoffload()?  ;-)

Ah it's called using work_on_cpu() which launch a workqueue on the
target and waits for completion. And that whole thing is protected
inside the barrier mutex and hotplug.

> Agreed!  And I do believe that concurrent callback execution will
> prove better than a possibly indefinite gap in callback execution.

Mutual agreement! :-)

Thanks.