Re: [RFC PATCH 48/86] rcu: handle quiescent states for PREEMPT_RCU=n

[Paul E. McKenney]

On Wed, Dec 06, 2023 at 04:10:18PM +0100, Thomas Gleixner wrote:
> Paul!
> 
> On Mon, Dec 04 2023 at 17:33, Paul E. McKenney wrote:
> > On Tue, Nov 28, 2023 at 06:04:33PM +0100, Thomas Gleixner wrote:
> >> So:
> >> 
> >>     loop()
> >> 
> >>       preempt_disable();
> >> 
> >>       --> tick interrupt
> >>             rcu_flavor_sched_clock_irq()
> >>                 sets NEED_RESCHED
> >> 
> >>       preempt_enable()
> >>         preempt_schedule()
> >>           schedule()
> >>             report_QS()
> >> 
> >> See? No magic nonsense in preempt_enable(), no cond_resched(), nothing.
> >
> > Understood, but that does delay detection of that quiescent state by up
> > to one tick.
> 
> Sure, but does that really matter in practice?

It might, but yes, I would expect it to matter far less than the other
things I have been calling out.

> >> So if that turns out to matter in reality and not just by academic
> >> inspection, then we are far better off to annotate such code with:
> >> 
> >>     do {
> >>         preempt_lazy_disable();
> >>         mutex_lock();
> >>         do_stuff();
> >>         mutex_unlock();
> >>         preempt_lazy_enable();
> >>     }
> >> 
> >> and let preempt_lazy_enable() evaluate the NEED_RESCHED_LAZY bit.
> >
> > I am not exactly sure what semantics you are proposing with this pairing
> > as opposed to "this would be a good time to preempt in response to the
> > pending lazy request".  But I do agree that something like this could
> > replace at least a few more instance of cond_resched(), so that is good.
> > Not necessarily all of them, though.
> 
> The main semantic difference is that such a mechanism is properly
> nesting and can be eventually subsumed into the actual locking
> constructs.

OK, fair enough.

And noting that testing should include workloads that exercise things
like mutex_lock() and mutex_trylock() fastpaths.

> >> Just insisting that RCU_PREEMPT=n requires cond_resched() and whatsoever
> >> is not really getting us anywhere.
> >
> > Except that this is not what is happening, Thomas.  ;-)
> >
> > You are asserting that all of the cond_resched() calls can safely be
> > eliminated.  That might well be, but more than assertion is required.
> > You have come up with some good ways of getting rid of some classes of
> > them, which is a very good and very welcome thing.  But that is not the
> > same as having proved that all of them may be safely removed.
> 
> Neither have you proven that any of them will be required with the new
> PREEMPT_LAZY model. :)

True.  But nor have you proven them unnecessary.  That will need to
wait for larger-scale testing.

> Your experience and knowledge in this area is certainly appreciated, but
> under the changed semantics of LAZY it's debatable whether observations
> and assumptions which are based on PREEMPT_NONE behaviour still apply.
> 
> We'll see.

That we will!

							Thanx, Paul