Re: [PATCH drm-misc-next 1/3] drm/sched: implement dynamic job flow control

[Boris Brezillon <boris.brezillon@collabora.com>]

On Wed, 27 Sep 2023 13:45:37 +0200
Danilo Krummrich <dakr@redhat.com> wrote:

> On 9/27/23 09:25, Boris Brezillon wrote:
> > On Wed, 27 Sep 2023 02:13:59 +0200
> > Danilo Krummrich <dakr@redhat.com> wrote:
> >   
> >> On 9/26/23 22:43, Luben Tuikov wrote:  
> >>> Hi,
> >>>
> >>> On 2023-09-24 18:43, Danilo Krummrich wrote:  
> >>>> Currently, job flow control is implemented simply by limiting the amount
> >>>> of jobs in flight. Therefore, a scheduler is initialized with a
> >>>> submission limit that corresponds to a certain amount of jobs.  
> >>>
> >>> "certain"? How about this instead:
> >>> " ... that corresponds to the number of jobs which can be sent
> >>>     to the hardware."?
> >>>      
> >>>>
> >>>> This implies that for each job drivers need to account for the maximum  
> >>>                                   ^,
> >>> Please add a comma after "job".
> >>>      
> >>>> job size possible in order to not overflow the ring buffer.  
> >>>
> >>> Well, different hardware designs would implement this differently.
> >>> Ideally, you only want pointers into the ring buffer, and then
> >>> the hardware consumes as much as it can. But this is a moot point
> >>> and it's always a good idea to have a "job size" hint from the client.
> >>> So this is a good patch.
> >>>
> >>> Ideally, you want to say that the hardware needs to be able to
> >>> accommodate the number of jobs which can fit in the hardware
> >>> queue times the largest job. This is a waste of resources
> >>> however, and it is better to give a hint as to the size of a job,
> >>> by the client. If the hardware can peek and understand dependencies,
> >>> on top of knowing the "size of the job", it can be an extremely
> >>> efficient scheduler.
> >>>      
> >>>>
> >>>> However, there are drivers, such as Nouveau, where the job size has a
> >>>> rather large range. For such drivers it can easily happen that job
> >>>> submissions not even filling the ring by 1% can block subsequent
> >>>> submissions, which, in the worst case, can lead to the ring run dry.
> >>>>
> >>>> In order to overcome this issue, allow for tracking the actual job size
> >>>> instead of the amount job jobs. Therefore, add a field to track a job's  
> >>>
> >>> "the amount job jobs." --> "the number of jobs."  
> >>
> >> Yeah, I somehow manage to always get this wrong, which I guess you noticed
> >> below already.
> >>
> >> That's all good points below - gonna address them.
> >>
> >> Did you see Boris' response regarding a separate callback in order to fetch
> >> the job's submission units dynamically? Since this is needed by PowerVR, I'd
> >> like to include this in V2. What's your take on that?
> >>
> >> My only concern with that would be that if I got what Boris was saying
> >> correctly calling
> >>
> >> WARN_ON(s_job->submission_units > sched->submission_limit);
> >>
> >> from drm_sched_can_queue() wouldn't work anymore, since this could indeed happen
> >> temporarily. I think this was also Christian's concern.  
> > 
> > Actually, I think that's fine to account for the max job size in the
> > first check, we're unlikely to have so many native fence waits that our
> > job can't fit in an empty ring buffer.
> >   
> 
> But it can happen, right? Hence, we can't have this check, do we?
> 

I theory, yes, in practice, given the size of the ring buffers, and the
size of a fence wait command, I guess we can refuse to queue a job (at
the driver level) if the maximum job size (static + maximum dynamic
part of the job) doesn't fit in the ring buffer. I that ever becomes a
problem, we can revisit it at that point.