Re: [PATCH 3/3] rv: Add explicit lockdep context for reactors

[Gabriele Monaco <gmonaco@redhat.com>]

On Tue, 2025-10-14 at 11:46 +0200, Thomas Weißschuh wrote:
> On Tue, Oct 14, 2025 at 09:38:09AM +0200, Nam Cao wrote:
> > The reactors are invoked in tracepoints' handlers, thus they must not
> > trigger another tracepoint, otherwise we may be stuck in an infinite loop.
> > (this is why preempt_enable_notrace() exists alongside preempt_enable()).
> 
> Sounds reasonable. However today not even the printk reactor satisfies this
> rule as it transitively calls trace_console().

That's a valid concern, I assume it would become a problem also if we wanted to
use locks inside event handlers, as it was discussed some time ago to better
handle concurrency.

> > I'm not familiar with the internal lockdep. But I think these would
> > trigger trace_lock_acquire() and trace_lock_release().
> 
> Indeed. Right now no monitor attaches to those tracepoints. We could
> prevent monitors from attaching to certain "well-known" tracepoints.
> But then we still need to manually track which those are, which is ugly.
> Or we move the invocation of the reactor to a workqueue/task_work.

I'm afraid also workqueues might open a rabbit-hole (waking up a task fights
with locks in many scheduling tracepoints).
At a quick glance task_works also do some IPI/wakeups that are traced.
If I get it correctly we are looking for something absolutely lock-free/trace-
free, I can't really think of much at the moment, maybe abusing RCU callbacks
but those would have their set of problems too.

As much as it might be interesting to write monitors on lockdep tracepoints,
this seems challenging.

We could opt for a foolproof Kconfig solution and prevent reactors if lockdep is
active (leaving only the error tracepoints that are hopefully still safe).

Gabriele