Re: [PATCH RFC 10/18] drm/scheduler: Add can_run_job callback

[Daniel Vetter <daniel@ffwll.ch>]

On Wed, Mar 08, 2023 at 04:19:17PM +0100, Karol Herbst wrote:
> On Wed, Mar 8, 2023 at 4:09 PM Christian König <christian.koenig@amd.com> wrote:
> >
> > Am 08.03.23 um 15:43 schrieb Karol Herbst:
> > > [SNIP]
> > > "further"? There was no discussion at all,
> >
> > Yeah, well that is exactly what I wanted to archive.
> >
> > >   you just started off like
> > > that. If you think somebody misses that connection, you can point out
> > > to documentation/videos whatever so the contributor can understand
> > > what's wrong with an approach. You did that, so that's fine. It's just
> > > starting off _any_ discussion with a "Well complete NAK" is terrible
> > > style. I'd feel uncomfortable if that happened to me and I'm sure
> > > there are enough people like that that we should be more reasonable
> > > with our replies. Just.. don't.
> > >
> > > We are all humans here and people react negatively to such things. And
> > > if people do it on purpose it just makes it worse.
> >
> > I completely see your point, I just don't know how to improve it.
> >
> > I don't stop people like this because I want to make them uncomfortable
> > but because I want to prevent further discussions on that topic.
> >
> > In other words how can I make people notice that this is something
> > fundamental while still being polite?

Ask them to improve the docs. Gets them on board, and for bonus point you
- can check they actually get it when you review the doc patch
- get scheduler docs for free
- have an easily pasteable link for next time around instead of just an
  aggressive NAK that helps no one really (aside from getting people
  boiling).

It's not really about being polite but making sure that efficient
communiction happens and that you don't have to repeat yourself. In rare
cases you get to type the docs themself when people are too dense to learn
(like what I had to do with the various dma_fence docs).

> I think a little improvement over this would be to at least wait a few
> replies before resorting to those strong statements. Just before it
> becomes a risk in just wasting time.

See above what I'm trying to do. When the message doesn't sink in as
either a proper doc patch or when linking to the doc patch for next time
around (because let's face it, this entire concept of "dma_fence committed
for execution" is extremely trick, there will be repeations of this
question until we've sunset dma_fence, which is probably decades away).

If the learning does not happen, then it's the time to whack the big
hammer (and if people don't get it, you can escalate to Dave&me, we have
tools to make sure people get the message). But this really should be the
end, not the start of the escalation chain :-)

Cheers, Daniel

> 
> > >>>> This is clearly going against the idea of having jobs only depend on
> > >>>> fences and nothing else which is mandatory for correct memory management.
> > >>>>
> > >>> I'm sure it's all documented and there is a design document on how
> > >>> things have to look like you can point out? Might help to get a better
> > >>> understanding on how things should be.
> > >> Yeah, that's the problematic part. We have documented this very
> > >> extensively:
> > >> https://www.kernel.org/doc/html/v5.9/driver-api/dma-buf.html#indefinite-dma-fences
> > >>
> > >> And both Jason and Daniel gave talks about the underlying problem and
> > > fyi:
> > > s/Jason/Faith/g
> >
> > +1. I wasn't aware of that.
> >
> > >> try to come up with patches to raise warnings when that happens, but
> > >> people still keep coming up with the same idea over and over again.
> > >>
> > > Yes, and we'll have to tell them over and over again. Nothing wrong
> > > with that. That's just part of maintaining such a big subsystem. And
> > > that's definitely not a valid reason to phrase things like above.
> > >
> > >> It's just that the technical relationship between preventing jobs from
> > >> running and with that preventing dma_fences from signaling and the core
> > >> memory management with page faults and shrinkers waiting for those
> > >> fences is absolutely not obvious.
> > >>
> > >> We had at least 10 different teams from different companies falling into
> > >> the same trap already and either the patches were rejected of hand or
> > >> had to painfully reverted or mitigated later on.
> > >>
> > > Sure, but that's just part of the job. And pointing out fundamental
> > > mistakes early on is important, but the situation won't get any better
> > > by being like that. Yes, we'll have to repeat the same words over and
> > > over again, and yes that might be annoying, but that's just how it is.
> >
> > Well I have no problem explaining people why a solution doesn't work.
> >
> > But what usually happens is that people don't realize that they need to
> > back of from a design and completely start over.
> >
> > Regards,
> > Christian.
> >
> > >
> > >> Regards,
> > >> Christian.
> > >>
> > >>>> If the hw is busy with something you need to return the fence for this
> > >>>> from the prepare_job callback so that the scheduler can be notified when
> > >>>> the hw is available again.
> > >>>>
> > >>>> Regards,
> > >>>> Christian.
> > >>>>
> > >>>>> Signed-off-by: Asahi Lina <lina@asahilina.net>
> > >>>>> ---
> > >>>>>     drivers/gpu/drm/scheduler/sched_main.c | 10 ++++++++++
> > >>>>>     include/drm/gpu_scheduler.h            |  8 ++++++++
> > >>>>>     2 files changed, 18 insertions(+)
> > >>>>>
> > >>>>> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> > >>>>> index 4e6ad6e122bc..5c0add2c7546 100644
> > >>>>> --- a/drivers/gpu/drm/scheduler/sched_main.c
> > >>>>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> > >>>>> @@ -1001,6 +1001,16 @@ static int drm_sched_main(void *param)
> > >>>>>                 if (!entity)
> > >>>>>                         continue;
> > >>>>>
> > >>>>> +             if (sched->ops->can_run_job) {
> > >>>>> +                     sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
> > >>>>> +                     if (!sched_job) {
> > >>>>> +                             complete_all(&entity->entity_idle);
> > >>>>> +                             continue;
> > >>>>> +                     }
> > >>>>> +                     if (!sched->ops->can_run_job(sched_job))
> > >>>>> +                             continue;
> > >>>>> +             }
> > >>>>> +
> > >>>>>                 sched_job = drm_sched_entity_pop_job(entity);
> > >>>>>
> > >>>>>                 if (!sched_job) {
> > >>>>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> > >>>>> index 9db9e5e504ee..bd89ea9507b9 100644
> > >>>>> --- a/include/drm/gpu_scheduler.h
> > >>>>> +++ b/include/drm/gpu_scheduler.h
> > >>>>> @@ -396,6 +396,14 @@ struct drm_sched_backend_ops {
> > >>>>>         struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job,
> > >>>>>                                          struct drm_sched_entity *s_entity);
> > >>>>>
> > >>>>> +     /**
> > >>>>> +      * @can_run_job: Called before job execution to check whether the
> > >>>>> +      * hardware is free enough to run the job.  This can be used to
> > >>>>> +      * implement more complex hardware resource policies than the
> > >>>>> +      * hw_submission limit.
> > >>>>> +      */
> > >>>>> +     bool (*can_run_job)(struct drm_sched_job *sched_job);
> > >>>>> +
> > >>>>>         /**
> > >>>>>              * @run_job: Called to execute the job once all of the dependencies
> > >>>>>              * have been resolved.  This may be called multiple times, if
> > >>>>>
> >
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch