On Mon, Jan 16, 2017 at 06:11:30PM +0100, Peter Zijlstra wrote:
> On Sat, Jan 14, 2017 at 01:13:12AM -0800, Paul E. McKenney wrote:
> > There is some confusion as to which of cond_resched() or
> > cond_resched_rcu_qs() should be added to long in-kernel loops.
> > This commit therefore eliminates the decision by adding RCU
> > quiescent states to cond_resched().
>
> Which would make: rcu_read_lock(); cond_resched(); rcu_read_unlock();
> invalid under preemptible RCU. Is it already?
In theory, yes. In practice, I just tested it with preemption and
lockdep enabled, and it didn't complain. If further testing finds
complaints, we can either fix those uses (preferred) or revert
this patch.
> > Warning: This is a prototype. For example, it does not correctly
> > handle Tasks RCU. Which is OK for the moment, given that no one
> > actually uses Tasks RCU yet.
>
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -4907,6 +4907,7 @@ int __sched _cond_resched(void)
> > preempt_schedule_common();
> > return 1;
> > }
> > + rcu_all_qs();
> > return 0;
> > }
>
> Still not a real fan of this, it does make cond_resched() touch a bunch
> more cachelines, also, I suppose that if we're going to do this, we
> should make __cond_resched_lock() and __cond_resched_softirq() act
> similarly.
Michal (now CCed) argues that having to distinguish between cond_resched()
and cond_resched_rcu_qs() is overly burdensome. Michal?
Any thoughts on how we might remove this burden without the additional
cache misses? I will take another look as well to see what could make
it lower cost. There are probably ways... Would it make sense to
have RCU maintain a need-rcu_all_qs() flage in the same cacheline as
the __preempt_count? Perhaps throttling the writes to this flag from
the RCU grace-period kthreads to once per 100 milliseconds or so?
Thanx, Paul
