On Fri, 2025-08-01 at 09:58 +0200, Nam Cao wrote:
> On Thu, Jul 31, 2025 at 09:47:10AM +0200, Gabriele Monaco wrote:
> > On Wed, 2025-07-30 at 14:45 +0200, Nam Cao wrote:
> > > Add "real-time scheduling" monitor, which validates that SCHED_RR
> > > and SCHED_FIFO tasks are scheduled before tasks with normal and
> > > extensible scheduling policies
> > 
> > Looks a very interesting monitor!
> > A few questions:
> > 
> > I assume this works with rt-throttle because it implies a dequeue,
> > right?
> > And you probably won't see that without explicit tracepoints..
> 
> It does work properly with rt-throttling:
>       root@yellow:~# ./rt-loop
>       [   74.357730] sched: RT throttling activated
>       [   74.357745] rv: rts: 0: violation detected
> 
> Looking at rt-throlling code, it does not dequeue tasks, only does
>       rt_rq->rt_throttled = 1;
>       rt_rq->rt_queued = 0;
> 
> so we are fine.

Wait, by /works properly/ you mean it reports a violation. I just
noticed you mention it in the description.

It's reasonable to request RT throttling disabled on sanely configured
RT systems. But throttling is a /valid/ kernel feature, I get you mark
it as /unwanted/ though.

I guess if that's the case, this monitor doesn't belong in the sched
collection because it's meant to validate the kernel behaviour, not its
configuration for a specific purpose (RT).
Isn't it better off with the rtapp ones (which validate the system
configuration to run in an RT scenario).

Does it make sense to you?

> > 
> > As far as I understand here the monitor would just miss RT tasks
> > already running but would perfectly enforce the ones starting after
> > initialisation, right?
> 
> Not exactly. What could happen is that:
> 
>  - RT task A already running
> 
>  - monitor enabled. The monitor isn't aware of task A, therefore it
> allows
>    sched_switch to switch to non-RT task.
> 
>  - RT task B is queued. The monitor now knows at least one RT task is
>    enqueued, so it disallows sched_switch to switch to non-RT.
> 
>  - RT task A is dequeued. However, the monitor does not differentiate
> task
>    A and task B, therefore I thinks the only enqueued RT task is now
> gone.
> 
>  - So now we have task B started after the monitor, but the monitor
> does not check it.
> 
> The monitor will become accurate once the CPU has no enqueued RT
> task, which should happen quite quickly on a sane setup where RT
> tasks do not monopoly the CPU.
> 
> The monitor could be changed to be accurate from the get-go, by
> looking at how many enqueued RT tasks are present. I *think* rt_rq-
> >rt_nr_running works. But I think the current implementation is
> fine, so not worth thinking too much about it.

Yeah if it's something quickly reached it shouldn't be a problem, also
rt throttle would run in case there's an RT monopoly and you'd see a
violation already.

> > 
> > Not sure you can do much about it though. (without falling into the
> > need resched rabbithole I was trying to untangle)
> 
> I would need to look into scheduler code, maybe we could check that
> the next scheduler tick implies a sched_switch. Another day.

Agree, the monitor looks good for now.
I still want to give it a run when I have a bit more time, besides with
RT throttle, can the monitor really fail on a working system?

Thanks,
Gabriele


Reply via email to