1 files changed, 357 insertions, 169 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
index 6614d8a6f4c8..468003583b2a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
@@ -26,6 +26,7 @@
 #include "amdgpu.h"
 #include "amdgpu_sched.h"
 #include "amdgpu_ras.h"
+#include <linux/nospec.h>
 
 #define to_amdgpu_ctx_entity(e)	\
 	container_of((e), struct amdgpu_ctx_entity, entity)
@@ -42,21 +43,61 @@ const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = {
 	[AMDGPU_HW_IP_VCN_JPEG]	=	1,
 };
 
-static int amdgpu_ctx_total_num_entities(void)
+bool amdgpu_ctx_priority_is_valid(int32_t ctx_prio)
 {
-	unsigned i, num_entities = 0;
+	switch (ctx_prio) {
+	case AMDGPU_CTX_PRIORITY_UNSET:
+	case AMDGPU_CTX_PRIORITY_VERY_LOW:
+	case AMDGPU_CTX_PRIORITY_LOW:
+	case AMDGPU_CTX_PRIORITY_NORMAL:
+	case AMDGPU_CTX_PRIORITY_HIGH:
+	case AMDGPU_CTX_PRIORITY_VERY_HIGH:
+		return true;
+	default:
+		return false;
+	}
+}
+
+static enum drm_sched_priority
+amdgpu_ctx_to_drm_sched_prio(int32_t ctx_prio)
+{
+	switch (ctx_prio) {
+	case AMDGPU_CTX_PRIORITY_UNSET:
+		return DRM_SCHED_PRIORITY_UNSET;
+
+	case AMDGPU_CTX_PRIORITY_VERY_LOW:
+		return DRM_SCHED_PRIORITY_MIN;
+
+	case AMDGPU_CTX_PRIORITY_LOW:
+		return DRM_SCHED_PRIORITY_MIN;
+
+	case AMDGPU_CTX_PRIORITY_NORMAL:
+		return DRM_SCHED_PRIORITY_NORMAL;
+
+	case AMDGPU_CTX_PRIORITY_HIGH:
+		return DRM_SCHED_PRIORITY_HIGH;
+
+	case AMDGPU_CTX_PRIORITY_VERY_HIGH:
+		return DRM_SCHED_PRIORITY_HIGH;
 
-	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i)
-		num_entities += amdgpu_ctx_num_entities[i];
+	/* This should not happen as we sanitized userspace provided priority
+	 * already, WARN if this happens.
+	 */
+	default:
+		WARN(1, "Invalid context priority %d\n", ctx_prio);
+		return DRM_SCHED_PRIORITY_NORMAL;
+	}
 
-	return num_entities;
 }
 
 static int amdgpu_ctx_priority_permit(struct drm_file *filp,
-				      enum drm_sched_priority priority)
+				      int32_t priority)
 {
+	if (!amdgpu_ctx_priority_is_valid(priority))
+		return -EINVAL;
+
 	/* NORMAL and below are accessible by everyone */
-	if (priority <= DRM_SCHED_PRIORITY_NORMAL)
+	if (priority <= AMDGPU_CTX_PRIORITY_NORMAL)
 		return 0;
 
 	if (capable(CAP_SYS_NICE))
@@ -68,47 +109,124 @@ static int amdgpu_ctx_priority_permit(struct drm_file *filp,
 	return -EACCES;
 }
 
+static enum amdgpu_gfx_pipe_priority amdgpu_ctx_prio_to_compute_prio(int32_t prio)
+{
+	switch (prio) {
+	case AMDGPU_CTX_PRIORITY_HIGH:
+	case AMDGPU_CTX_PRIORITY_VERY_HIGH:
+		return AMDGPU_GFX_PIPE_PRIO_HIGH;
+	default:
+		return AMDGPU_GFX_PIPE_PRIO_NORMAL;
+	}
+}
+
+static enum amdgpu_ring_priority_level amdgpu_ctx_sched_prio_to_ring_prio(int32_t prio)
+{
+	switch (prio) {
+	case AMDGPU_CTX_PRIORITY_HIGH:
+		return AMDGPU_RING_PRIO_1;
+	case AMDGPU_CTX_PRIORITY_VERY_HIGH:
+		return AMDGPU_RING_PRIO_2;
+	default:
+		return AMDGPU_RING_PRIO_0;
+	}
+}
+
+static unsigned int amdgpu_ctx_get_hw_prio(struct amdgpu_ctx *ctx, u32 hw_ip)
+{
+	struct amdgpu_device *adev = ctx->adev;
+	int32_t ctx_prio;
+	unsigned int hw_prio;
+
+	ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
+			ctx->init_priority : ctx->override_priority;
+
+	switch (hw_ip) {
+	case AMDGPU_HW_IP_COMPUTE:
+		hw_prio = amdgpu_ctx_prio_to_compute_prio(ctx_prio);
+		break;
+	case AMDGPU_HW_IP_VCE:
+	case AMDGPU_HW_IP_VCN_ENC:
+		hw_prio = amdgpu_ctx_sched_prio_to_ring_prio(ctx_prio);
+		break;
+	default:
+		hw_prio = AMDGPU_RING_PRIO_DEFAULT;
+		break;
+	}
+
+	hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
+	if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0)
+		hw_prio = AMDGPU_RING_PRIO_DEFAULT;
+
+	return hw_prio;
+}
+
+
+static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
+				  const u32 ring)
+{
+	struct amdgpu_device *adev = ctx->adev;
+	struct amdgpu_ctx_entity *entity;
+	struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
+	unsigned num_scheds = 0;
+	int32_t ctx_prio;
+	unsigned int hw_prio;
+	enum drm_sched_priority drm_prio;
+	int r;
+
+	entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs),
+			 GFP_KERNEL);
+	if (!entity)
+		return  -ENOMEM;
+
+	ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
+			ctx->init_priority : ctx->override_priority;
+	entity->sequence = 1;
+	hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip);
+	drm_prio = amdgpu_ctx_to_drm_sched_prio(ctx_prio);
+
+	hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
+	scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
+	num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
+
+	/* 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 = drm_sched_pick_best(scheds, num_scheds);
+		scheds = &sched;
+		num_scheds = 1;
+	}
+
+	r = drm_sched_entity_init(&entity->entity, drm_prio, scheds, num_scheds,
+				  &ctx->guilty);
+	if (r)
+		goto error_free_entity;
+
+	ctx->entities[hw_ip][ring] = entity;
+	return 0;
+
+error_free_entity:
+	kfree(entity);
+
+	return r;
+}
+
 static int amdgpu_ctx_init(struct amdgpu_device *adev,
-			   enum drm_sched_priority priority,
+			   int32_t priority,
 			   struct drm_file *filp,
 			   struct amdgpu_ctx *ctx)
 {
-	unsigned num_entities = amdgpu_ctx_total_num_entities();
-	unsigned i, j, k;
 	int r;
 
-	if (priority < 0 || priority >= DRM_SCHED_PRIORITY_MAX)
-		return -EINVAL;
-
 	r = amdgpu_ctx_priority_permit(filp, priority);
 	if (r)
 		return r;
 
 	memset(ctx, 0, sizeof(*ctx));
-	ctx->adev = adev;
 
-	ctx->fences = kcalloc(amdgpu_sched_jobs * num_entities,
-			      sizeof(struct dma_fence*), GFP_KERNEL);
-	if (!ctx->fences)
-		return -ENOMEM;
-
-	ctx->entities[0] = kcalloc(num_entities,
-				   sizeof(struct amdgpu_ctx_entity),
-				   GFP_KERNEL);
-	if (!ctx->entities[0]) {
-		r = -ENOMEM;
-		goto error_free_fences;
-	}
-
-	for (i = 0; i < num_entities; ++i) {
-		struct amdgpu_ctx_entity *entity = &ctx->entities[0][i];
-
-		entity->sequence = 1;
-		entity->fences = &ctx->fences[amdgpu_sched_jobs * i];
-	}
-	for (i = 1; i < AMDGPU_HW_IP_NUM; ++i)
-		ctx->entities[i] = ctx->entities[i - 1] +
-			amdgpu_ctx_num_entities[i - 1];
+	ctx->adev = adev;
 
 	kref_init(&ctx->refcount);
 	spin_lock_init(&ctx->ring_lock);
@@ -118,116 +236,50 @@ static int amdgpu_ctx_init(struct amdgpu_device *adev,
 	ctx->reset_counter_query = ctx->reset_counter;
 	ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
 	ctx->init_priority = priority;
-	ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
+	ctx->override_priority = AMDGPU_CTX_PRIORITY_UNSET;
 
-	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
-		struct amdgpu_ring *rings[AMDGPU_MAX_RINGS];
-		struct drm_sched_rq *rqs[AMDGPU_MAX_RINGS];
-		unsigned num_rings = 0;
-		unsigned num_rqs = 0;
-
-		switch (i) {
-		case AMDGPU_HW_IP_GFX:
-			rings[0] = &adev->gfx.gfx_ring[0];
-			num_rings = 1;
-			break;
-		case AMDGPU_HW_IP_COMPUTE:
-			for (j = 0; j < adev->gfx.num_compute_rings; ++j)
-				rings[j] = &adev->gfx.compute_ring[j];
-			num_rings = adev->gfx.num_compute_rings;
-			break;
-		case AMDGPU_HW_IP_DMA:
-			for (j = 0; j < adev->sdma.num_instances; ++j)
-				rings[j] = &adev->sdma.instance[j].ring;
-			num_rings = adev->sdma.num_instances;
-			break;
-		case AMDGPU_HW_IP_UVD:
-			rings[0] = &adev->uvd.inst[0].ring;
-			num_rings = 1;
-			break;
-		case AMDGPU_HW_IP_VCE:
-			rings[0] = &adev->vce.ring[0];
-			num_rings = 1;
-			break;
-		case AMDGPU_HW_IP_UVD_ENC:
-			rings[0] = &adev->uvd.inst[0].ring_enc[0];
-			num_rings = 1;
-			break;
-		case AMDGPU_HW_IP_VCN_DEC:
-			for (j = 0; j < adev->vcn.num_vcn_inst; ++j) {
-				if (adev->vcn.harvest_config & (1 << j))
-					continue;
-				rings[num_rings++] = &adev->vcn.inst[j].ring_dec;
-			}
-			break;
-		case AMDGPU_HW_IP_VCN_ENC:
-			for (j = 0; j < adev->vcn.num_vcn_inst; ++j) {
-				if (adev->vcn.harvest_config & (1 << j))
-					continue;
-				for (k = 0; k < adev->vcn.num_enc_rings; ++k)
-					rings[num_rings++] = &adev->vcn.inst[j].ring_enc[k];
-			}
-			break;
-		case AMDGPU_HW_IP_VCN_JPEG:
-			for (j = 0; j < adev->vcn.num_vcn_inst; ++j) {
-				if (adev->vcn.harvest_config & (1 << j))
-					continue;
-				rings[num_rings++] = &adev->vcn.inst[j].ring_jpeg;
-			}
-			break;
-		}
-
-		for (j = 0; j < num_rings; ++j) {
-			if (!rings[j]->adev)
-				continue;
+	return 0;
+}
 
-			rqs[num_rqs++] = &rings[j]->sched.sched_rq[priority];
-		}
+static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
+{
 
-		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j)
-			r = drm_sched_entity_init(&ctx->entities[i][j].entity,
-						  rqs, num_rqs, &ctx->guilty);
-		if (r)
-			goto error_cleanup_entities;
-	}
+	int i;
 
-	return 0;
+	if (!entity)
+		return;
 
-error_cleanup_entities:
-	for (i = 0; i < num_entities; ++i)
-		drm_sched_entity_destroy(&ctx->entities[0][i].entity);
-	kfree(ctx->entities[0]);
+	for (i = 0; i < amdgpu_sched_jobs; ++i)
+		dma_fence_put(entity->fences[i]);
 
-error_free_fences:
-	kfree(ctx->fences);
-	ctx->fences = NULL;
-	return r;
+	kfree(entity);
 }
 
 static void amdgpu_ctx_fini(struct kref *ref)
 {
 	struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount);
-	unsigned num_entities = amdgpu_ctx_total_num_entities();
 	struct amdgpu_device *adev = ctx->adev;
 	unsigned i, j;
 
 	if (!adev)
 		return;
 
-	for (i = 0; i < num_entities; ++i)
-		for (j = 0; j < amdgpu_sched_jobs; ++j)
-			dma_fence_put(ctx->entities[0][i].fences[j]);
-	kfree(ctx->fences);
-	kfree(ctx->entities[0]);
+	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
+		for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) {
+			amdgpu_ctx_fini_entity(ctx->entities[i][j]);
+			ctx->entities[i][j] = NULL;
+		}
+	}
 
 	mutex_destroy(&ctx->lock);
-
 	kfree(ctx);
 }
 
 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
 			  u32 ring, struct drm_sched_entity **entity)
 {
+	int r;
+
 	if (hw_ip >= AMDGPU_HW_IP_NUM) {
 		DRM_ERROR("unknown HW IP type: %d\n", hw_ip);
 		return -EINVAL;
@@ -244,14 +296,20 @@ int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
 		return -EINVAL;
 	}
 
-	*entity = &ctx->entities[hw_ip][ring].entity;
+	if (ctx->entities[hw_ip][ring] == NULL) {
+		r = amdgpu_ctx_init_entity(ctx, hw_ip, ring);
+		if (r)
+			return r;
+	}
+
+	*entity = &ctx->entities[hw_ip][ring]->entity;
 	return 0;
 }
 
 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
 			    struct amdgpu_fpriv *fpriv,
 			    struct drm_file *filp,
-			    enum drm_sched_priority priority,
+			    int32_t priority,
 			    uint32_t *id)
 {
 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
@@ -284,14 +342,17 @@ static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
 static void amdgpu_ctx_do_release(struct kref *ref)
 {
 	struct amdgpu_ctx *ctx;
-	unsigned num_entities;
-	u32 i;
+	u32 i, j;
 
 	ctx = container_of(ref, struct amdgpu_ctx, refcount);
+	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
+		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
+			if (!ctx->entities[i][j])
+				continue;
 
-	num_entities = amdgpu_ctx_total_num_entities();
-	for (i = 0; i < num_entities; i++)
-		drm_sched_entity_destroy(&ctx->entities[0][i].entity);
+			drm_sched_entity_destroy(&ctx->entities[i][j]->entity);
+		}
+	}
 
 	amdgpu_ctx_fini(ref);
 }
@@ -345,13 +406,15 @@ static int amdgpu_ctx_query(struct amdgpu_device *adev,
 	return 0;
 }
 
+#define AMDGPU_RAS_COUNTE_DELAY_MS 3000
+
 static int amdgpu_ctx_query2(struct amdgpu_device *adev,
-	struct amdgpu_fpriv *fpriv, uint32_t id,
-	union drm_amdgpu_ctx_out *out)
+			     struct amdgpu_fpriv *fpriv, uint32_t id,
+			     union drm_amdgpu_ctx_out *out)
 {
+	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 	struct amdgpu_ctx *ctx;
 	struct amdgpu_ctx_mgr *mgr;
-	unsigned long ras_counter;
 
 	if (!fpriv)
 		return -EINVAL;
@@ -376,19 +439,28 @@ static int amdgpu_ctx_query2(struct amdgpu_device *adev,
 	if (atomic_read(&ctx->guilty))
 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY;
 
-	/*query ue count*/
-	ras_counter = amdgpu_ras_query_error_count(adev, false);
-	/*ras counter is monotonic increasing*/
-	if (ras_counter != ctx->ras_counter_ue) {
-		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE;
-		ctx->ras_counter_ue = ras_counter;
-	}
+	if (adev->ras_enabled && con) {
+		/* Return the cached values in O(1),
+		 * and schedule delayed work to cache
+		 * new vaues.
+		 */
+		int ce_count, ue_count;
+
+		ce_count = atomic_read(&con->ras_ce_count);
+		ue_count = atomic_read(&con->ras_ue_count);
+
+		if (ce_count != ctx->ras_counter_ce) {
+			ctx->ras_counter_ce = ce_count;
+			out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE;
+		}
+
+		if (ue_count != ctx->ras_counter_ue) {
+			ctx->ras_counter_ue = ue_count;
+			out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE;
+		}
 
-	/*query ce count*/
-	ras_counter = amdgpu_ras_query_error_count(adev, true);
-	if (ras_counter != ctx->ras_counter_ce) {
-		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE;
-		ctx->ras_counter_ce = ras_counter;
+		schedule_delayed_work(&con->ras_counte_delay_work,
+				      msecs_to_jiffies(AMDGPU_RAS_COUNTE_DELAY_MS));
 	}
 
 	mutex_unlock(&mgr->lock);
@@ -400,20 +472,19 @@ int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
 {
 	int r;
 	uint32_t id;
-	enum drm_sched_priority priority;
+	int32_t priority;
 
 	union drm_amdgpu_ctx *args = data;
-	struct amdgpu_device *adev = dev->dev_private;
+	struct amdgpu_device *adev = drm_to_adev(dev);
 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
 
-	r = 0;
 	id = args->in.ctx_id;
-	priority = amdgpu_to_sched_priority(args->in.priority);
+	priority = args->in.priority;
 
 	/* For backwards compatibility reasons, we need to accept
 	 * ioctls with garbage in the priority field */
-	if (priority == DRM_SCHED_PRIORITY_INVALID)
-		priority = DRM_SCHED_PRIORITY_NORMAL;
+	if (!amdgpu_ctx_priority_is_valid(priority))
+		priority = AMDGPU_CTX_PRIORITY_NORMAL;
 
 	switch (args->in.op) {
 	case AMDGPU_CTX_OP_ALLOC_CTX:
@@ -465,7 +536,7 @@ int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
 
 void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
 			  struct drm_sched_entity *entity,
-			  struct dma_fence *fence, uint64_t* handle)
+			  struct dma_fence *fence, uint64_t *handle)
 {
 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
 	uint64_t seq = centity->sequence;
@@ -518,22 +589,49 @@ struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 	return fence;
 }
 
+static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
+					   struct amdgpu_ctx_entity *aentity,
+					   int hw_ip,
+					   int32_t priority)
+{
+	struct amdgpu_device *adev = ctx->adev;
+	unsigned int hw_prio;
+	struct drm_gpu_scheduler **scheds = NULL;
+	unsigned num_scheds;
+
+	/* set sw priority */
+	drm_sched_entity_set_priority(&aentity->entity,
+				      amdgpu_ctx_to_drm_sched_prio(priority));
+
+	/* set hw priority */
+	if (hw_ip == AMDGPU_HW_IP_COMPUTE) {
+		hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip);
+		hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX);
+		scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
+		num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
+		drm_sched_entity_modify_sched(&aentity->entity, scheds,
+					      num_scheds);
+	}
+}
+
 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
-				  enum drm_sched_priority priority)
+				  int32_t priority)
 {
-	unsigned num_entities = amdgpu_ctx_total_num_entities();
-	enum drm_sched_priority ctx_prio;
-	unsigned i;
+	int32_t ctx_prio;
+	unsigned i, j;
 
 	ctx->override_priority = priority;
 
-	ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
+	ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
 			ctx->init_priority : ctx->override_priority;
+	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
+		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
+			if (!ctx->entities[i][j])
+				continue;
 
-	for (i = 0; i < num_entities; i++) {
-		struct drm_sched_entity *entity = &ctx->entities[0][i].entity;
-
-		drm_sched_entity_set_priority(entity, ctx_prio);
+			amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j],
+						       i, ctx_prio);
+		}
 	}
 }
 
@@ -569,20 +667,24 @@ void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
 
 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
 {
-	unsigned num_entities = amdgpu_ctx_total_num_entities();
 	struct amdgpu_ctx *ctx;
 	struct idr *idp;
-	uint32_t id, i;
+	uint32_t id, i, j;
 
 	idp = &mgr->ctx_handles;
 
 	mutex_lock(&mgr->lock);
 	idr_for_each_entry(idp, ctx, id) {
-		for (i = 0; i < num_entities; i++) {
-			struct drm_sched_entity *entity;
+		for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
+			for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
+				struct drm_sched_entity *entity;
+
+				if (!ctx->entities[i][j])
+					continue;
 
-			entity = &ctx->entities[0][i].entity;
-			timeout = drm_sched_entity_flush(entity, timeout);
+				entity = &ctx->entities[i][j]->entity;
+				timeout = drm_sched_entity_flush(entity, timeout);
+			}
 		}
 	}
 	mutex_unlock(&mgr->lock);
@@ -591,10 +693,9 @@ long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
 
 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
 {
-	unsigned num_entities = amdgpu_ctx_total_num_entities();
 	struct amdgpu_ctx *ctx;
 	struct idr *idp;
-	uint32_t id, i;
+	uint32_t id, i, j;
 
 	idp = &mgr->ctx_handles;
 
@@ -604,8 +705,17 @@ void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
 			continue;
 		}
 
-		for (i = 0; i < num_entities; i++)
-			drm_sched_entity_fini(&ctx->entities[0][i].entity);
+		for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
+			for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
+				struct drm_sched_entity *entity;
+
+				if (!ctx->entities[i][j])
+					continue;
+
+				entity = &ctx->entities[i][j]->entity;
+				drm_sched_entity_fini(entity);
+			}
+		}
 	}
 }
 
@@ -627,3 +737,81 @@ void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
 	idr_destroy(&mgr->ctx_handles);
 	mutex_destroy(&mgr->lock);
 }
+
+static void amdgpu_ctx_fence_time(struct amdgpu_ctx *ctx,
+		struct amdgpu_ctx_entity *centity, ktime_t *total, ktime_t *max)
+{
+	ktime_t now, t1;
+	uint32_t i;
+
+	*total = *max = 0;
+
+	now = ktime_get();
+	for (i = 0; i < amdgpu_sched_jobs; i++) {
+		struct dma_fence *fence;
+		struct drm_sched_fence *s_fence;
+
+		spin_lock(&ctx->ring_lock);
+		fence = dma_fence_get(centity->fences[i]);
+		spin_unlock(&ctx->ring_lock);
+		if (!fence)
+			continue;
+		s_fence = to_drm_sched_fence(fence);
+		if (!dma_fence_is_signaled(&s_fence->scheduled)) {
+			dma_fence_put(fence);
+			continue;
+		}
+		t1 = s_fence->scheduled.timestamp;
+		if (!ktime_before(t1, now)) {
+			dma_fence_put(fence);
+			continue;
+		}
+		if (dma_fence_is_signaled(&s_fence->finished) &&
+			s_fence->finished.timestamp < now)
+			*total += ktime_sub(s_fence->finished.timestamp, t1);
+		else
+			*total += ktime_sub(now, t1);
+		t1 = ktime_sub(now, t1);
+		dma_fence_put(fence);
+		*max = max(t1, *max);
+	}
+}
+
+ktime_t amdgpu_ctx_mgr_fence_usage(struct amdgpu_ctx_mgr *mgr, uint32_t hwip,
+		uint32_t idx, uint64_t *elapsed)
+{
+	struct idr *idp;
+	struct amdgpu_ctx *ctx;
+	uint32_t id;
+	struct amdgpu_ctx_entity *centity;
+	ktime_t total = 0, max = 0;
+
+	if (idx >= AMDGPU_MAX_ENTITY_NUM)
+		return 0;
+	idp = &mgr->ctx_handles;
+	mutex_lock(&mgr->lock);
+	idr_for_each_entry(idp, ctx, id) {
+		ktime_t ttotal, tmax;
+
+		if (!ctx->entities[hwip][idx])
+			continue;
+
+		centity = ctx->entities[hwip][idx];
+		amdgpu_ctx_fence_time(ctx, centity, &ttotal, &tmax);
+
+		/* Harmonic mean approximation diverges for very small
+		 * values. If ratio < 0.01% ignore
+		 */
+		if (AMDGPU_CTX_FENCE_USAGE_MIN_RATIO(tmax, ttotal))
+			continue;
+
+		total = ktime_add(total, ttotal);
+		max = ktime_after(tmax, max) ? tmax : max;
+	}
+
+	mutex_unlock(&mgr->lock);
+	if (elapsed)
+		*elapsed = max;
+
+	return total;
+}