On Thu, 21 Jul 2022 14:08:38 +0200
Daniel Bristot de Oliveira <bristot@xxxxxxxxxx> wrote:
> On 7/20/22 22:06, Steven Rostedt wrote:
> >> +/* \
> >> + * da_monitor_enabled_##name - checks if the monitor is enabled \
> >> + */ \
> >> +static inline bool da_monitor_enabled_##name(void) \
> >> +{ \
> > Should we add a:
> >
> > smp_rmb();
> >
> > here? And then a smp_wmb() where these switches get updated?
> >
>
> Makes sense.
>
> Should I also add the READ_ONCE/WRITE_ONCE? like
>
> smp_rmb()
> READ_ONCE(var)
>
> WRITE_ONCE(var, value)
> smp_wmb()
I'm not sure the WRITE_ONCE() is necessary with the memory barriers.
Because they should also prevent gcc from doing anything after that
barrier. As Linus once stated, most cases WRITE_ONCE() is useless, but it's
fine to keep more for annotation (as to pair with the READ_ONCE()) than for
anything that is critical.
>
> for all these on/off knobs, or just the barriers?
>
> > I guess how critical is it that these turn off immediately after the switch
> > is flipped?
>
> It is not critical to continue the execution of those that have already crossed by
> the variable. Still, waiting for the tracepoints to finish their execution before
> returning to the user-space task that disabled the variable might be a good thing.
You mean after disabling, to wait for the tracepoints that are currently
running to end?
>
> IIRC, we can do that via RCU... like, synchronize_rcu()?
We have tracepoint_synchronize_unregister() that does that, as some
traceponits use SRCU and not RCU.
-- Steve
>
> >> + /* global switch */ \
> >> + if (unlikely(!rv_monitoring_on())) \
> >> + return 0; \
> >> + \
> >> + /* monitor enabled */ \
> >> + if (unlikely(!rv_##name.enabled)) \
> >> + return 0; \
> >> + \
> >> + return 1; \
> >> +} \
> >> + \
