Re: [PATCH RT v2 2/3] sched: migrate_enable: Use sleeping_lock to indicate involuntary sleep

From: Paul E. McKenney
Date: Tue Aug 27 2019 - 11:58:24 EST

On Tue, Aug 27, 2019 at 09:08:53AM -0400, Joel Fernandes wrote:
> On Tue, Aug 27, 2019 at 11:23:33AM +0200, Sebastian Andrzej Siewior wrote:
> [snip]
> > > However, if this was instead an rcu_read_lock() critical section within
> > > a PREEMPT=y kernel, then if a schedule() occured within stop_one_task(),
> > > RCU would consider that critical section to be preempted. This means
> > > that any RCU grace period that is blocked by this RCU read-side critical
> > > section would remain blocked until stop_one_cpu() resumed, returned,
> > > and so on until the matching rcu_read_unlock() was reached. In other
> > > words, RCU would consider that RCU read-side critical section to span
> > > the call to stop_one_cpu() even if stop_one_cpu() invoked schedule().
> >
> > Isn't that my example from above and what we do in RT? My understanding
> > is that this is the reason why we need BOOST on RT otherwise the RCU
> > critical section could remain blocked for some time.
> Not just for boost, it is needed to block the grace period itself on
> PREEMPT=y. On PREEMPT=y, if rcu_note_context_switch() happens in middle of a
> rcu_read_lock() reader section, then the task is added to a blocked list
> (rcu_preempt_ctxt_queue). Then just after that, the CPU reports a QS state
> (rcu_qs()) as you can see in the PREEMPT=y implementation of
> rcu_note_context_switch(). Even though the CPU has reported a QS, the grace
> period will not end because the preempted (or block as could be in -rt) task
> is still blocking the grace period. This is fundamental to the function of
> Preemptible-RCU where there is the concept of tasks blocking a grace period,
> not just CPUs.
> I think what Paul is trying to explain AIUI (Paul please let me know if I
> missed something):
> (1) Anyone calling rcu_note_context_switch() and expecting it to respect
> RCU-readers that are readers as a result of interrupt disabled regions, have
> incorrect expectations. So calling rcu_note_context_switch() has to be done
> carefully.
> (2) Disabling interrupts is "generally" implied as an RCU-sched flavor
> reader. However, invoking rcu_note_context_switch() from a disabled interrupt
> region is *required* for rcu_note_context_switch() to work correctly.
> (3) On PREEMPT=y kernels, invoking rcu_note_context_switch() from an
> interrupt disabled region does not mean that that the task will be added to a
> blocked list (unless it is also in an RCU-preempt reader) so
> rcu_note_context_switch() may immediately report a quiescent state and
> nothing blockings the grace period.
> So callers of rcu_note_context_switch() must be aware of this behavior.
> (4) On PREEMPT=n, unlike PREEMPT=y, there is no blocked list handling and so
> nothing will block the grace period once rcu_note_context_switch() is called.
> So any path calling rcu_note_context_switch() on a PREEMPT=n kernel, in the
> middle of something that is expected to be an RCU reader would be really bad
> from an RCU view point.
> Probably, we should add this all to documentation somewhere.

I think that Sebastian understands this and was using the example of RCU
priority boosting to confirm his understanding. But documentation would
be good. Extremely difficult to keep current, but good. I believe that
the requirements document does cover this.

Thanx, Paul

> thanks!
> - Joel
> > > On the other hand, within a PREEMPT=n kernel, the call to schedule()
> > > would split even an rcu_read_lock() critical section. Which is why I
> > > asked earlier if sleeping_lock_inc() and sleeping_lock_dec() are no-ops
> > > in !PREEMPT_RT_BASE kernels. We would after all want the usual lockdep
> > > complaints in that case.
> >
> > sleeping_lock_inc() +dec() is only RT specific. It is part of RT's
> > spin_lock() implementation and used by RCU (rcu_note_context_switch())
> > to not complain if invoked within a critical section.
> >
> > > Does that help, or am I missing the point?
> > >
> > > Thanx, Paul
> > Sebastian