On Fri, May 16, 2025 at 12:35:52PM +0100, Tvrtko Ursulin wrote:
>
> On 16/05/2025 11:54, Danilo Krummrich wrote:
> > On Fri, May 16, 2025 at 11:19:50AM +0100, Tvrtko Ursulin wrote:
> > >
> > > On 16/05/2025 10:53, Danilo Krummrich wrote:
> > > > On Fri, May 16, 2025 at 10:33:30AM +0100, Tvrtko Ursulin wrote:
> > > > > On 24/04/2025 10:55, Philipp Stanner wrote:
> > > > > > + * @kill_fence_context: kill the fence context belonging to
> > > > > > this scheduler
> > > > >
> > > > > Which fence context would that be? ;)
> > > >
> > > > There's one one per ring and a scheduler instance represents a single
> > > > ring. So,
> > > > what should be specified here?
> > >
> > > I was pointing out the fact not all drivers are 1:1 sched:entity.
> >
> > I'm well aware, but how is that relevant? Entities don't have an associated
> > fence context, but a GPU Ring (either hardware or software) has, which a
> > scheduler instance represents.
>
> Aha! Well.. how it is relevant and do entities not have an associated fence
> context? Well, entity->fence_context.. that was my first association this
> whole time. Never it crossed my mind this is talking about the hardware
> fence context. Proof in the pudding naming should be improved.
It says "fence context belonging to this scheduler", which should be
unambiguous, however I agree that we could mark out the difference even more.
> But I also don't think there is a requirement for fences returned from
> ->run_job() to have a single context. Which again makes it not the best
> naming.
It's implied by the fact that a scheduler instance represents a ring. Having
multiple fence contexts per ring doesn't make any sense.
But it's indeed not written down -- we should do that then.
> > In the callback the driver should neither tear down an entity, nor the whole
> > scheduler, hence we shouldn't call it like that. sched_kill() is therefore
> > misleading as well.
>
> ->sched_exit()? ->sched_stop()? ->sched_cleanup()?
I think this all would throw up questions like "What does {exit,stop,cleanup}
mean in this context?". And the answer would be "kill the fence context of the
ring represented by the scheduler".
I think we want a name that represents that without an indirection that we have
to define.
> > It should be named after what it actually does (or should do). Feel free to
> > propose a different name that conforms with that.
> >
> > > > > We also probably want some commentary on the topic of indefinite (or
> > > > > very
> > > > > long at least) blocking a thread exit / SIGINT/TERM/KILL time.
> > > >
> > > > You mean in case the driver does implement the callback, but does *not*
> > > > properly
> > > > tear down the fence context? So, you ask for describing potential
> > > > consequences
> > > > of drivers having bugs in the implementation of the callback? Or
> > > > something else?
> > >
> > > I was proposing the kerneldoc for the vfunc should document the callback
> > > must not block, or if blocking is unavoidable, either document a guideline
> > > on how long is acceptable. Maybe even enforce a limit in the scheduler
> > > core
> > > itself.
> >
> > Killing the fence context shouldn't block.
>
> Cool. And maybe convert the wait_event to wait_event_timeout with a warning
> to be robust.
That would make sense if it could deadlock, but even if the driver does nothing
it should terminate eventually. The rule that we always rely on is that we
guarantee throughout the kernel that fences are signalled eventually.
