Re: [PATCH] drm/scheduler: fix last_scheduled handling

Ping, Nayan any comments on that or can I commit it?

This is just a stripped down version of my original last_scheduled improvement patch.

Christian.

Am 07.08.2018 um 14:54 schrieb Christian König:

Make sure we access last_scheduled only after checking that there are no more jobs on the entity.

Signed-off-by: Christian König christian.koenig@amd.com

drivers/gpu/drm/scheduler/gpu_scheduler.c | 21 +++++++++++---------- 1 file changed, 11 insertions(+), 10 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/gpu_scheduler.c b/drivers/gpu/drm/scheduler/gpu_scheduler.c index 8ee249886473..bd7883d1b964 100644 --- a/drivers/gpu/drm/scheduler/gpu_scheduler.c +++ b/drivers/gpu/drm/scheduler/gpu_scheduler.c @@ -568,19 +568,20 @@ void drm_sched_entity_push_job(struct drm_sched_job *sched_job, struct drm_sched_entity *entity) { struct drm_sched_rq *rq = entity->rq;

• bool first, reschedule, idle;

• bool first;

• idle = entity->last_scheduled == NULL ||

• dma_fence_is_signaled(entity->last_scheduled);
  

  first = spsc_queue_count(&entity->job_queue) == 0;
• reschedule = idle && first && (entity->num_rq_list > 1);

• if (first && (entity->num_rq_list > 1)) {

• struct dma_fence *fence;
  
  

• if (reschedule) {

• rq = drm_sched_entity_get_free_sched(entity);
  

• spin_lock(&entity->rq_lock);
  

• drm_sched_rq_remove_entity(entity->rq, entity);
  

• entity->rq = rq;
  

• spin_unlock(&entity->rq_lock);
  

• fence = READ_ONCE(entity->last_scheduled);
  

• if (fence == NULL || dma_fence_is_signaled(fence)) {
  

• 	rq = drm_sched_entity_get_free_sched(entity);
  

• 	spin_lock(&entity->rq_lock);
  

• 	drm_sched_rq_remove_entity(entity->rq, entity);
  

• 	entity->rq = rq;
  

• 	spin_unlock(&entity->rq_lock);
  

• }
  

  }

  sched_job->sched = entity->rq->sched;