[RFC PATCH 2/2] drm/amdgpu: disable gpu load balancer for vcn jobs

[Christian König]

Am 11.03.20 um 18:18 schrieb Nirmoy Das:
> VCN HW  doesn't support dynamic load balance on multiple
> instances for a context. This modifies the entity's sched_list
> to a sched_list consist of only one drm scheduler.
>
> Signed-off-by: Nirmoy Das <nirmoy.das at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c  |  1 +
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 25 +++++++++++++++++++++++++
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h |  2 ++
>   drivers/gpu/drm/amd/amdgpu/amdgpu_job.c |  1 +
>   4 files changed, 29 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> index 8304d0c87899..00032093d8a9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> @@ -1257,6 +1257,7 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
>   	priority = job->base.s_priority;
>   	drm_sched_entity_push_job(&job->base, entity);
>   
> +	amdgpu_ctx_limit_load_balance(entity);
>   	amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
>   
>   	ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> index fa575bdc03c8..57b49188306d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> @@ -139,6 +139,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, const u32 hw_ip, const
>   	if (r)
>   		goto error_free_entity;
>   
> +	entity->hw_ip = hw_ip;
>   	ctx->entities[hw_ip][ring] = entity;
>   	return 0;
>   
> @@ -559,6 +560,30 @@ void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
>   	}
>   }
>   
> +static void limit_vcn_load_balance(struct amdgpu_ctx_entity *centity)
> +{
> +	struct drm_gpu_scheduler **scheds = &centity->entity.rq->sched;
> +
> +	if (drm_sched_entity_num_jobs(&centity->entity) == 1)

That check doesn't work correctly, the job might actually already be 
processed when we hit here.

> +		drm_sched_entity_modify_sched(&centity->entity, scheds, 1);

Just always update the scheduler here.

> +
> +}
> +
> +void amdgpu_ctx_limit_load_balance(struct drm_sched_entity *entity)
> +{
> +	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
> +
> +	if (!centity)
> +		return;

That check looks superfluous to me.

> +
> +	switch (centity->hw_ip) {

Better get the ring from entity->rq->sched instead.

> +	case AMDGPU_HW_IP_VCN_DEC:
> +	case AMDGPU_HW_IP_VCN_ENC:

Maybe better to make that a flag in the ring functions, but this way 
works as well.

Regards,
Christian.

> +		limit_vcn_load_balance(centity);
> +	}
> +
> +}
> +
>   int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
>   			       struct drm_sched_entity *entity)
>   {
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h
> index de490f183af2..d52d8d562d77 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h
> @@ -33,6 +33,7 @@ struct amdgpu_fpriv;
>   
>   struct amdgpu_ctx_entity {
>   	uint64_t		sequence;
> +	uint32_t		hw_ip;
>   	struct drm_sched_entity	entity;
>   	struct dma_fence	*fences[];
>   };
> @@ -90,5 +91,6 @@ void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr);
>   
>   void amdgpu_ctx_init_sched(struct amdgpu_device *adev);
>   
> +void amdgpu_ctx_limit_load_balance(struct drm_sched_entity *entity);
>   
>   #endif
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> index 4981e443a884..955d12bc89ae 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> @@ -154,6 +154,7 @@ int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,
>   	amdgpu_job_free_resources(job);
>   	priority = job->base.s_priority;
>   	drm_sched_entity_push_job(&job->base, entity);
> +	amdgpu_ctx_limit_load_balance(entity);
>   
>   	return 0;
>   }