Hi Christian,
On Mon, Aug 6, 2018 at 5:49 PM Christian König <ckoenig.leichtzumerken@xxxxxxxxx> wrote:
Update job earlier with the scheduler it is supposed to be scheduled on.
Otherwise we could incorrectly optimize dependencies when moving an
entity from one scheduler to another.
I don't think change is really required, the old code should work fine. Read below for more comments.
Signed-off-by: Christian König <christian.koenig@xxxxxxx>
---
drivers/gpu/drm/scheduler/gpu_scheduler.c | 4 ++--
drivers/gpu/drm/scheduler/sched_fence.c | 2 +-
2 files changed, 3 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/scheduler/gpu_scheduler.c b/drivers/gpu/drm/scheduler/gpu_scheduler.c
index 65078dd3c82c..029863726c99 100644
--- a/drivers/gpu/drm/scheduler/gpu_scheduler.c
+++ b/drivers/gpu/drm/scheduler/gpu_scheduler.c
@@ -533,8 +533,6 @@ drm_sched_entity_pop_job(struct drm_sched_entity *entity)
if (!sched_job)
return NULL;
- sched_job->sched = sched;
- sched_job->s_fence->sched = sched;
while ((entity->dependency = sched->ops->dependency(sched_job, entity)))
The job should have the correct scheduler before the sched->oops->dependency() function is called for the first time. Hence I placed the assignment here.
It might also be that I am missing some case so please let me know if such a case exists where this might lead to wrong optimization.
Regards,
Nayan
if (drm_sched_entity_add_dependency_cb(entity))
return NULL;
@@ -581,6 +579,8 @@ void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
spin_unlock(&entity->rq_lock);
}
+ sched_job->sched = entity->rq->sched;
+ sched_job->s_fence->sched = entity->rq->sched;
trace_drm_sched_job(sched_job, entity);
atomic_inc(&entity->rq->sched->num_jobs);
WRITE_ONCE(entity->last_user, current->group_leader);
diff --git a/drivers/gpu/drm/scheduler/sched_fence.c b/drivers/gpu/drm/scheduler/sched_fence.c
index 4029312fdd81..6dab18d288d7 100644
--- a/drivers/gpu/drm/scheduler/sched_fence.c
+++ b/drivers/gpu/drm/scheduler/sched_fence.c
@@ -172,7 +172,7 @@ struct drm_sched_fence *drm_sched_fence_create(struct drm_sched_entity *entity,
return NULL;
fence->owner = owner;
- fence->sched = entity->rq->sched;
+ fence->sched = NULL;
spin_lock_init(&fence->lock);
seq = atomic_inc_return(&entity->fence_seq);
--
2.14.1
_______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel