[PATCH v2 2/3] drm/amdgpu: increment sched score on entity selection

[Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>]

For hw engines that can't load balance jobs, entities are
"statically" load balanced: on their first submit, they select
the best scheduler based on its score.
The score is made up of 2 parts:
* the job queue depth (how much jobs are executing/waiting)
* the number of entities assigned

The second part is only relevant for the static load balance:
it's a way to consider how many entities are attached to this
scheduler, knowing that if they ever submit jobs they will go
to this one.

For rings that can load balance jobs freely, idle entities
aren't a concern and shouldn't impact the scheduler's decisions.

Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 23 +++++++++++++++++------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h |  1 +
 2 files changed, 18 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
index f5d5c45ddc0d..953c81c928c1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
@@ -209,6 +209,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
 	struct amdgpu_ctx_entity *entity;
 	enum drm_sched_priority drm_prio;
 	unsigned int hw_prio, num_scheds;
+	struct amdgpu_ring *aring;
 	int32_t ctx_prio;
 	int r;
 
@@ -236,14 +237,16 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
 		r = amdgpu_xcp_select_scheds(adev, hw_ip, hw_prio, fpriv,
 						&num_scheds, &scheds);
 		if (r)
-			goto cleanup_entity;
+			goto error_free_entity;
 	}
 
-	/* disable load balance if the hw engine retains context among dependent jobs */
-	if (hw_ip == AMDGPU_HW_IP_VCN_ENC ||
-	    hw_ip == AMDGPU_HW_IP_VCN_DEC ||
-	    hw_ip == AMDGPU_HW_IP_UVD_ENC ||
-	    hw_ip == AMDGPU_HW_IP_UVD) {
+	sched = scheds[0];
+	aring = container_of(sched, struct amdgpu_ring, sched);
+
+	if (aring->funcs->engine_retains_context) {
+		/* Disable load balancing between multiple schedulers if the hw
+		 * engine retains context among dependent jobs.
+		 */
 		sched = drm_sched_pick_best(scheds, num_scheds);
 		scheds = &sched;
 		num_scheds = 1;
@@ -258,6 +261,11 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
 	if (cmpxchg(&ctx->entities[hw_ip][ring], NULL, entity))
 		goto cleanup_entity;
 
+	if (aring->funcs->engine_retains_context) {
+		entity->sched_score = sched->score;
+		atomic_inc(entity->sched_score);
+	}
+
 	return 0;
 
 cleanup_entity:
@@ -514,6 +522,9 @@ static void amdgpu_ctx_do_release(struct kref *ref)
 			if (!ctx->entities[i][j])
 				continue;
 
+			if (ctx->entities[i][j]->sched_score)
+				atomic_dec(ctx->entities[i][j]->sched_score);
+
 			drm_sched_entity_destroy(&ctx->entities[i][j]->entity);
 		}
 	}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h
index 090dfe86f75b..f7b44f96f374 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h
@@ -39,6 +39,7 @@ struct amdgpu_ctx_entity {
 	uint32_t		hw_ip;
 	uint64_t		sequence;
 	struct drm_sched_entity	entity;
+	atomic_t		*sched_score;
 	struct dma_fence	*fences[];
 };
 
-- 
2.43.0