On Mon, Jul 09, 2018 at 09:34:32AM -0700, Paul E. McKenney wrote:
> On Mon, Jul 09, 2018 at 05:26:32PM +0200, Peter Zijlstra wrote:
> > On Mon, Jul 09, 2018 at 07:29:32AM -0700, Paul E. McKenney wrote:
> > > OK, so here are our options:
> > >
> > > 1. Add the RCU conditional to need_resched(), as David suggests.
> > > Peter has concerns about overhead.
> > >
> > > 2. Create a new need_resched_rcu_qs() that is to be used when
> > > deciding whether or not to do cond_resched(). This would
> > > exact the overhead only where it is needed, but is one more
> > > thing for people to get wrong.
> >
> > Also, with the crypto guys checking need_resched() in asm that won't
> > really work either.
>
> Fair point! Ease of use is a good thing, even within the Linux kernel.
> Or maybe especially within the Linux kernel...
>
> > > 3. Revert my changes to de-emphasize cond_resched_rcu_qs(),
> > > and go back to sprinkling cond_resched_rcu_qs() throughout
> > > the code. This also is one more thing for people to get wrong,
> > > and might well eventually convert all cond_resched() calls to
> > > cond_resched_rcu_qs(), which sure seems like a failure mode to me.
> >
> > 4a. use resched_cpu() more agressive
> > 4b. use the tick to set TIF_NEED_RESCHED when it finds rcu_urgent_qs
> > (avoids the IPI at the 'cost' of a slight delay in processing)
>
> 4b sounds eminently reasonable to me! Something like the (untested,
> probably doesn't even build) patch below?
>
> David, any reason why this wouldn't work? Seems to me that this would
> make need_resched() respond to RCU's need for quiescent states in a
> timely manner without need_resched() having to become heavier weight,
> but figured I should ask.
>
> > 5. make guest mode a quiescent state (like supposedly already done
> > for NOHZ_FULL) (but this would not help the crypto guys).
> >
> > 6. ....
> >
> > ok I ran out of ideas here I think.
> >
> >
> > So for PREEMPT the tick can check preempt_count() == 0 and if so, know
> > it _could_ have rescheduled and advance the qs, right? But since we
> > don't have a preempt count for !PREEMPT_COUNT kernels this doesn't work.
> >
> > And thus we need to invoke actual scheduling events and then through the
> > schedule() callback RCU knows things happened.
> >
> > 4b seems like something worth trying for !PREEMPT kernels I suppose
>
> David is running a !PREEMPT kernel.
>
> For PREEMPT kernels, the patch below results in a quiescent state for
> the CPU, and the forced schedule queues the task. This queuing enables
> later RCU priority boosting (if enabled) once all other CPUs sharing
> the same leaf rcu_node structure have passed through quiescent states.
>
> And yes, for PREEMPT kernels the scheduling-clock interrupt handler
> already checks for a quiescent state using a combination of
> preempt_count() (as you say, but ignoring the hardirq bits because
> we are in an interrupt handler) and current->rcu_read_lock_nesting.
>
> So I believe that this will cover it.
>
> Thoughts?
Updated per Peter's feedback on IRC.
Thanx, Paul
------------------------------------------------------------------------
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 51919985f6cf..ccde5f8aff61 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2496,6 +2496,12 @@ void rcu_check_callbacks(int user)
{
trace_rcu_utilization(TPS("Start scheduler-tick"));
raw_cpu_inc(rcu_data.ticks_this_gp);
+ if (smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)) &&
+ !is_idle_task(current)) {
+ set_tsk_need_resched(current);
+ set_preempt_need_resched();
+ }
+ __this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
rcu_flavor_check_callbacks(user);
if (rcu_pending())
invoke_rcu_core();
