On 2018-03-12 14:27:29 [+0100], Peter Zijlstra wrote: > On Mon, Mar 12, 2018 at 11:51:13AM +0100, Sebastian Andrzej Siewior wrote: > > On 2018-03-09 23:26:43 [+0100], Peter Zijlstra wrote: > > > On Fri, Mar 09, 2018 at 09:25:50PM +0100, Sebastian Andrzej Siewior wrote: > > > > Is it just about the irqsave() usage or something else? I doubt it is > > > > the list walk. It is still unbound if not called from irq-off region. > > > > > > The current list walk is preemptible. You put the entire iteration (of > > > unbound length) inside a single critical section which destroy RT. > > > > I considered that list walk as cheap. We don't do any wake ups with the > > list walk - just mark the task for a later wake up. But if it is not I > > could add an upper limit of 20 iterations or so. > > So the problem is that as soon as this is exposed to userspace you've > lost. I know. We had this very same problem with clock_nanosleep() which got solved after timer rework. > If a user can stack like 10000 tasks on the completion before triggering > it, you've got yourself a giant !preempt section. Yes the wake_q stuff > is cheaper, but unbound is still unbound. > > wake_all must not be used from !preemptible (IRQ or otherwise) sections. > And I'm not seeing how waking just the top 20 helps. I assumed you complained about the unbounded list-walk with interrupts disabled (which is cheap but unbound is unbound). So here I suggested I move 20 entries off that list a time and enable interrupts again so an interrupt could set need_resched. But if we get invoked !preemptible then nothing changes. > > > Why isn't this a problem on RT? > > So we remain in the preempt_disable() section due to RCU-sched so we > > have this, yes. But the "disabled interrupts" part is due to > > spin_lock_irqsave() which is a non-issue on RT. So if we managed to get > > rid of the rcu-sched then the swait can go and we can stick with the > > wake_up_all() on RT, too. > > OK, so for RT we simply loose the IRQ-disable thing, but its still a > !preemptible section. exactly. The irqsafe() was to guard non-RT config which uses the same code. So do I understand you correctly that irqsafe may remain for !RT config but that invocation with disabled preemption due to sched_rcu (on RT, too) must go? Sebastian -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
