On Mon, Aug 07, 2023 at 05:46:16PM +0200, Christian König wrote:
> Am 04.08.23 um 16:13 schrieb Matthew Brost:
> > [SNIP]
> > Christian / Daniel - I've read both of you comments and having a hard
> > time parsing them. I do not really understand the issue with this patch
> > or exactly what is being suggested instead. Let's try to work through
> > this.
> >
> > > > > > I'm still extremely frowned on this.
> > > > > >
> > > > > > If you need this functionality then let the drivers decide which
> > > > > > runqueue the scheduler should use.
> > What do you mean by runqueue here? Do you mean 'struct
> > workqueue_struct'? The scheduler in this context is 'struct
> > drm_gpu_scheduler', right?
>
> Sorry for the confusing wording, your understanding is correct.
>
> > Yes, we have added this functionality iin the first patch.
> >
> > > > > > When you then create a single threaded runqueue you can just submit work
> > > > > > to it and serialize this with the scheduler work.
> > > > > >
> > We don't want to use a single threaded workqueue_struct in Xe, we want
> > to use a system_wq as run_job() can be executed in parallel across
> > multiple entites (or drm_gpu_scheduler as in Xe we have 1 to 1
> > relationship between entity and scheduler). What we want is on per
> > entity / scheduler granularity to be able to communicate into the
> > backend a message synchronously (run_job / free_job not executing,
> > scheduler execution not paused for a reset).
> >
> > If I'm underatanding what you suggesting in Xe we'd create an ordered
> > workqueue_struct per drm_gpu_scheduler and the queue messages on the
> > ordered workqueue_struct?
>
> Yes, correct.
>
> > This seems pretty messy to me as now we have
> > open coded a solution bypassing the scheduler, every drm_gpu_scheduler
> > creates its own workqueue_struct, and we'd also have to open code the
> > pausing of these messages for resets too.
> >
> > IMO this is pretty clean solution that follows the pattern of cleanup
> > jobs already in place.
>
> Yeah, exactly that's the point. Moving the job cleanup into the scheduler
> thread is seen as very very bad idea by me.
>
> And I really don't want to exercise that again for different use cases.
>
> >
> > > > > > This way we wouldn't duplicate this core kernel function inside the
> > > > > > scheduler.
> > > > > Yeah that's essentially the design we picked for the tdr workers,
> > > > > where some drivers have requirements that all tdr work must be done on
> > > > > the same thread (because of cross-engine coordination issues). But
> > > > > that would require that we rework the scheduler as a pile of
> > > > > self-submitting work items, and I'm not sure that actually fits all
> > > > > that well into the core workqueue interfaces either.
> > This is the ordering between TDRs firing between different
> > drm_gpu_scheduler and larger external resets (GT in Xe) an ordered
> > workqueue_struct makes sense for this. Here we are talking about
> > ordering jobs and messages within a single drm_gpu_scheduler. Using the
> > main execution thread to do ordering makes sense in my opinion.
>
> I completely disagree to that.
>
> Take a look at how this came to be. This is a very very ugly hack and we
> already had a hard time making lockdep understand the different fence
> signaling dependencies with freeing the job and I'm pretty sure that is
> still not 100% correct.
>
> >
> > > > There were already patches floating around which did exactly that.
> > > >
> > > > Last time I checked those were actually looking pretty good.
> > > >
> > Link to patches for reference.
> >
> > > > Additional to message passing advantage the real big issue with the
> > > > scheduler and 1 to 1 mapping is that we create a kernel thread for each
> > > > instance, which results in tons on overhead.
> > First patch in the series switches from kthread to work queue, that is
> > still a good idea.
>
> This was the patch I was referring to. Sorry didn't remembered that this was
> in the same patch set.
>
> >
> > > > Just using a work item which is submitted to a work queue completely avoids
> > > > that.
> > > Hm I should have read the entire series first, since that does the
> > > conversion still. Apologies for the confusion, and yeah we should be able
> > > to just submit other work to the same wq with the first patch? And so
> > > hand-rolling this infra here isn't needed at all?
> > >
> > I wouldn't call this hand rolling, rather it following patten already in
> > place.
>
> Basically workqueues are the in kernel infrastructure for exactly that use
> case and we are trying to re-create that here and that is usually a rather
> bad idea.
>
Ok let me play around with what this would look like in Xe, what you are
suggesting would be ordered-wq per scheduler, work item for run job,
work item for clean up job, and work item for a message. That might
work I suppose? Only issue I see is scaling as this exposes an
ordered-wq creation directly to an IOCTL. No idea if that is actually a
concern though.
Matt
> Regards,
> Christian.
>
> >
> > Matt
> >
> > > Or what am I missing?
> > >
> > > > Regards,
> > > > Christian.
> > > >
> > > > > Worst case I think this isn't a dead-end and can be refactored to
> > > > > internally use the workqueue services, with the new functions here
> > > > > just being dumb wrappers until everyone is converted over. So it
> > > > > doesn't look like an expensive mistake, if it turns out to be a
> > > > > mistake.
> > > > > -Daniel
> > > > >
> > > > >
> > > > > > Regards,
> > > > > > Christian.
> > > > > >
> > > > > > > Signed-off-by: Matthew Brost <matthew.brost@xxxxxxxxx>
> > > > > > > ---
> > > > > > > drivers/gpu/drm/scheduler/sched_main.c | 52 +++++++++++++++++++++++++-
> > > > > > > include/drm/gpu_scheduler.h | 29 +++++++++++++-
> > > > > > > 2 files changed, 78 insertions(+), 3 deletions(-)
> > > > > > >
> > > > > > > diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> > > > > > > index 2597fb298733..84821a124ca2 100644
> > > > > > > --- a/drivers/gpu/drm/scheduler/sched_main.c
> > > > > > > +++ b/drivers/gpu/drm/scheduler/sched_main.c
> > > > > > > @@ -1049,6 +1049,49 @@ drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
> > > > > > > }
> > > > > > > EXPORT_SYMBOL(drm_sched_pick_best);
> > > > > > >
> > > > > > > +/**
> > > > > > > + * drm_sched_add_msg - add scheduler message
> > > > > > > + *
> > > > > > > + * @sched: scheduler instance
> > > > > > > + * @msg: message to be added
> > > > > > > + *
> > > > > > > + * Can and will pass an jobs waiting on dependencies or in a runnable queue.
> > > > > > > + * Messages processing will stop if schedule run wq is stopped and resume when
> > > > > > > + * run wq is started.
> > > > > > > + */
> > > > > > > +void drm_sched_add_msg(struct drm_gpu_scheduler *sched,
> > > > > > > + struct drm_sched_msg *msg)
> > > > > > > +{
> > > > > > > + spin_lock(&sched->job_list_lock);
> > > > > > > + list_add_tail(&msg->link, &sched->msgs);
> > > > > > > + spin_unlock(&sched->job_list_lock);
> > > > > > > +
> > > > > > > + drm_sched_run_wq_queue(sched);
> > > > > > > +}
> > > > > > > +EXPORT_SYMBOL(drm_sched_add_msg);
> > > > > > > +
> > > > > > > +/**
> > > > > > > + * drm_sched_get_msg - get scheduler message
> > > > > > > + *
> > > > > > > + * @sched: scheduler instance
> > > > > > > + *
> > > > > > > + * Returns NULL or message
> > > > > > > + */
> > > > > > > +static struct drm_sched_msg *
> > > > > > > +drm_sched_get_msg(struct drm_gpu_scheduler *sched)
> > > > > > > +{
> > > > > > > + struct drm_sched_msg *msg;
> > > > > > > +
> > > > > > > + spin_lock(&sched->job_list_lock);
> > > > > > > + msg = list_first_entry_or_null(&sched->msgs,
> > > > > > > + struct drm_sched_msg, link);
> > > > > > > + if (msg)
> > > > > > > + list_del(&msg->link);
> > > > > > > + spin_unlock(&sched->job_list_lock);
> > > > > > > +
> > > > > > > + return msg;
> > > > > > > +}
> > > > > > > +
> > > > > > > /**
> > > > > > > * drm_sched_main - main scheduler thread
> > > > > > > *
> > > > > > > @@ -1060,6 +1103,7 @@ static void drm_sched_main(struct work_struct *w)
> > > > > > > container_of(w, struct drm_gpu_scheduler, work_run);
> > > > > > > struct drm_sched_entity *entity;
> > > > > > > struct drm_sched_job *cleanup_job;
> > > > > > > + struct drm_sched_msg *msg;
> > > > > > > int r;
> > > > > > >
> > > > > > > if (READ_ONCE(sched->pause_run_wq))
> > > > > > > @@ -1067,12 +1111,15 @@ static void drm_sched_main(struct work_struct *w)
> > > > > > >
> > > > > > > cleanup_job = drm_sched_get_cleanup_job(sched);
> > > > > > > entity = drm_sched_select_entity(sched);
> > > > > > > + msg = drm_sched_get_msg(sched);
> > > > > > >
> > > > > > > - if (!entity && !cleanup_job)
> > > > > > > + if (!entity && !cleanup_job && !msg)
> > > > > > > return; /* No more work */
> > > > > > >
> > > > > > > if (cleanup_job)
> > > > > > > sched->ops->free_job(cleanup_job);
> > > > > > > + if (msg)
> > > > > > > + sched->ops->process_msg(msg);
> > > > > > >
> > > > > > > if (entity) {
> > > > > > > struct dma_fence *fence;
> > > > > > > @@ -1082,7 +1129,7 @@ static void drm_sched_main(struct work_struct *w)
> > > > > > > sched_job = drm_sched_entity_pop_job(entity);
> > > > > > > if (!sched_job) {
> > > > > > > complete_all(&entity->entity_idle);
> > > > > > > - if (!cleanup_job)
> > > > > > > + if (!cleanup_job && !msg)
> > > > > > > return; /* No more work */
> > > > > > > goto again;
> > > > > > > }
> > > > > > > @@ -1177,6 +1224,7 @@ int drm_sched_init(struct drm_gpu_scheduler *sched,
> > > > > > >
> > > > > > > init_waitqueue_head(&sched->job_scheduled);
> > > > > > > INIT_LIST_HEAD(&sched->pending_list);
> > > > > > > + INIT_LIST_HEAD(&sched->msgs);
> > > > > > > spin_lock_init(&sched->job_list_lock);
> > > > > > > atomic_set(&sched->hw_rq_count, 0);
> > > > > > > INIT_DELAYED_WORK(&sched->work_tdr, drm_sched_job_timedout);
> > > > > > > diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> > > > > > > index df1993dd44ae..267bd060d178 100644
> > > > > > > --- a/include/drm/gpu_scheduler.h
> > > > > > > +++ b/include/drm/gpu_scheduler.h
> > > > > > > @@ -394,6 +394,23 @@ enum drm_gpu_sched_stat {
> > > > > > > DRM_GPU_SCHED_STAT_ENODEV,
> > > > > > > };
> > > > > > >
> > > > > > > +/**
> > > > > > > + * struct drm_sched_msg - an in-band (relative to GPU scheduler run queue)
> > > > > > > + * message
> > > > > > > + *
> > > > > > > + * Generic enough for backend defined messages, backend can expand if needed.
> > > > > > > + */
> > > > > > > +struct drm_sched_msg {
> > > > > > > + /** @link: list link into the gpu scheduler list of messages */
> > > > > > > + struct list_head link;
> > > > > > > + /**
> > > > > > > + * @private_data: opaque pointer to message private data (backend defined)
> > > > > > > + */
> > > > > > > + void *private_data;
> > > > > > > + /** @opcode: opcode of message (backend defined) */
> > > > > > > + unsigned int opcode;
> > > > > > > +};
> > > > > > > +
> > > > > > > /**
> > > > > > > * struct drm_sched_backend_ops - Define the backend operations
> > > > > > > * called by the scheduler
> > > > > > > @@ -471,6 +488,12 @@ struct drm_sched_backend_ops {
> > > > > > > * and it's time to clean it up.
> > > > > > > */
> > > > > > > void (*free_job)(struct drm_sched_job *sched_job);
> > > > > > > +
> > > > > > > + /**
> > > > > > > + * @process_msg: Process a message. Allowed to block, it is this
> > > > > > > + * function's responsibility to free message if dynamically allocated.
> > > > > > > + */
> > > > > > > + void (*process_msg)(struct drm_sched_msg *msg);
> > > > > > > };
> > > > > > >
> > > > > > > /**
> > > > > > > @@ -482,6 +505,7 @@ struct drm_sched_backend_ops {
> > > > > > > * @timeout: the time after which a job is removed from the scheduler.
> > > > > > > * @name: name of the ring for which this scheduler is being used.
> > > > > > > * @sched_rq: priority wise array of run queues.
> > > > > > > + * @msgs: list of messages to be processed in @work_run
> > > > > > > * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
> > > > > > > * waits on this wait queue until all the scheduled jobs are
> > > > > > > * finished.
> > > > > > > @@ -489,7 +513,7 @@ struct drm_sched_backend_ops {
> > > > > > > * @job_id_count: used to assign unique id to the each job.
> > > > > > > * @run_wq: workqueue used to queue @work_run
> > > > > > > * @timeout_wq: workqueue used to queue @work_tdr
> > > > > > > - * @work_run: schedules jobs and cleans up entities
> > > > > > > + * @work_run: schedules jobs, cleans up jobs, and processes messages
> > > > > > > * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
> > > > > > > * timeout interval is over.
> > > > > > > * @pending_list: the list of jobs which are currently in the job queue.
> > > > > > > @@ -513,6 +537,7 @@ struct drm_gpu_scheduler {
> > > > > > > long timeout;
> > > > > > > const char *name;
> > > > > > > struct drm_sched_rq sched_rq[DRM_SCHED_PRIORITY_COUNT];
> > > > > > > + struct list_head msgs;
> > > > > > > wait_queue_head_t job_scheduled;
> > > > > > > atomic_t hw_rq_count;
> > > > > > > atomic64_t job_id_count;
> > > > > > > @@ -566,6 +591,8 @@ void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
> > > > > > >
> > > > > > > void drm_sched_job_cleanup(struct drm_sched_job *job);
> > > > > > > void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
> > > > > > > +void drm_sched_add_msg(struct drm_gpu_scheduler *sched,
> > > > > > > + struct drm_sched_msg *msg);
> > > > > > > void drm_sched_run_wq_stop(struct drm_gpu_scheduler *sched);
> > > > > > > void drm_sched_run_wq_start(struct drm_gpu_scheduler *sched);
> > > > > > > void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
> > > --
> > > Daniel Vetter
> > > Software Engineer, Intel Corporation
> > > http://blog.ffwll.ch
>
