Merge branch 'drm-next-4.3' of git://people.freedesktop.org/~agd5f/linux into drm-next

- DP fixes for radeon and amdgpu
- IH ring fix for tonga and fiji
- Lots of GPU scheduler fixes
- Misc additional fixes

* 'drm-next-4.3' of git://people.freedesktop.org/~agd5f/linux: (42 commits)
  drm/amdgpu: fix wait queue handling in the scheduler
  drm/amdgpu: remove extra parameters from scheduler callbacks
  drm/amdgpu: wake up scheduler only when neccessary
  drm/amdgpu: remove entity idle timeout v2
  drm/amdgpu: fix postclose order
  drm/amdgpu: use IB for copy buffer of eviction
  drm/amdgpu: adjust the judgement of removing fence callback
  drm/amdgpu: fix no sync_wait in copy_buffer
  drm/amdgpu: fix last_vm_update fence is not effetive for sched fence
  drm/amdgpu: add priv data to sched
  drm/amdgpu: add owner for sched fence
  drm/amdgpu: remove entity reference from sched fence
  drm/amdgpu: fix and cleanup amd_sched_entity_push_job
  drm/amdgpu: remove amdgpu_bo_list_clone
  drm/amdgpu: remove the context from amdgpu_job
  drm/amdgpu: remove unused parameters to amd_sched_create
  drm/amdgpu: remove sched_lock
  drm/amdgpu: remove prepare_job callback
  drm/amdgpu: cleanup a scheduler function name
  drm/amdgpu: reorder scheduler functions
  ...

30 files changed, 671 insertions, 727 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 2fc58e658986..aa2dcf578dd6 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -183,6 +183,7 @@ struct amdgpu_vm;
 struct amdgpu_ring;
 struct amdgpu_semaphore;
 struct amdgpu_cs_parser;
+struct amdgpu_job;
 struct amdgpu_irq_src;
 struct amdgpu_fpriv;
 
@@ -246,7 +247,7 @@ struct amdgpu_buffer_funcs {
 	unsigned	copy_num_dw;
 
 	/* used for buffer migration */
-	void (*emit_copy_buffer)(struct amdgpu_ring *ring,
+	void (*emit_copy_buffer)(struct amdgpu_ib *ib,
 				 /* src addr in bytes */
 				 uint64_t src_offset,
 				 /* dst addr in bytes */
@@ -439,9 +440,12 @@ int amdgpu_fence_wait_next(struct amdgpu_ring *ring);
 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring);
 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring);
 
-signed long amdgpu_fence_wait_any(struct amdgpu_device *adev,
-			  struct amdgpu_fence **fences,
-			  bool intr, long t);
+signed long amdgpu_fence_wait_multiple(struct amdgpu_device *adev,
+				       struct fence **array,
+				       uint32_t count,
+				       bool wait_all,
+				       bool intr,
+				       signed long t);
 struct amdgpu_fence *amdgpu_fence_ref(struct amdgpu_fence *fence);
 void amdgpu_fence_unref(struct amdgpu_fence **fence);
 
@@ -514,7 +518,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring,
 		       uint64_t dst_offset,
 		       uint32_t byte_count,
 		       struct reservation_object *resv,
-		       struct amdgpu_fence **fence);
+		       struct fence **fence);
 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma);
 
 struct amdgpu_bo_list_entry {
@@ -650,7 +654,7 @@ struct amdgpu_sa_bo {
 	struct amdgpu_sa_manager	*manager;
 	unsigned			soffset;
 	unsigned			eoffset;
-	struct amdgpu_fence		*fence;
+	struct fence		        *fence;
 };
 
 /*
@@ -692,7 +696,7 @@ bool amdgpu_semaphore_emit_wait(struct amdgpu_ring *ring,
 				struct amdgpu_semaphore *semaphore);
 void amdgpu_semaphore_free(struct amdgpu_device *adev,
 			   struct amdgpu_semaphore **semaphore,
-			   struct amdgpu_fence *fence);
+			   struct fence *fence);
 
 /*
  * Synchronization
@@ -700,7 +704,8 @@ void amdgpu_semaphore_free(struct amdgpu_device *adev,
 struct amdgpu_sync {
 	struct amdgpu_semaphore *semaphores[AMDGPU_NUM_SYNCS];
 	struct amdgpu_fence	*sync_to[AMDGPU_MAX_RINGS];
-	struct amdgpu_fence	*last_vm_update;
+	DECLARE_HASHTABLE(fences, 4);
+	struct fence	        *last_vm_update;
 };
 
 void amdgpu_sync_create(struct amdgpu_sync *sync);
@@ -712,8 +717,9 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
 		     void *owner);
 int amdgpu_sync_rings(struct amdgpu_sync *sync,
 		      struct amdgpu_ring *ring);
+int amdgpu_sync_wait(struct amdgpu_sync *sync);
 void amdgpu_sync_free(struct amdgpu_device *adev, struct amdgpu_sync *sync,
-		      struct amdgpu_fence *fence);
+		      struct fence *fence);
 
 /*
  * GART structures, functions & helpers
@@ -871,7 +877,7 @@ int amdgpu_sched_ib_submit_kernel_helper(struct amdgpu_device *adev,
 					 struct amdgpu_ring *ring,
 					 struct amdgpu_ib *ibs,
 					 unsigned num_ibs,
-					 int (*free_job)(struct amdgpu_cs_parser *),
+					 int (*free_job)(struct amdgpu_job *),
 					 void *owner,
 					 struct fence **fence);
 
@@ -957,7 +963,7 @@ struct amdgpu_vm_id {
 	unsigned		id;
 	uint64_t		pd_gpu_addr;
 	/* last flushed PD/PT update */
-	struct amdgpu_fence	*flushed_updates;
+	struct fence	        *flushed_updates;
 	/* last use of vmid */
 	struct amdgpu_fence	*last_id_use;
 };
@@ -1042,7 +1048,7 @@ struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id);
 int amdgpu_ctx_put(struct amdgpu_ctx *ctx);
 
 uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
-			      struct fence *fence, uint64_t queued_seq);
+			      struct fence *fence);
 struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 				   struct amdgpu_ring *ring, uint64_t seq);
 
@@ -1078,8 +1084,6 @@ struct amdgpu_bo_list {
 };
 
 struct amdgpu_bo_list *
-amdgpu_bo_list_clone(struct amdgpu_bo_list *list);
-struct amdgpu_bo_list *
 amdgpu_bo_list_get(struct amdgpu_fpriv *fpriv, int id);
 void amdgpu_bo_list_put(struct amdgpu_bo_list *list);
 void amdgpu_bo_list_free(struct amdgpu_bo_list *list);
@@ -1255,14 +1259,16 @@ struct amdgpu_cs_parser {
 
 	/* user fence */
 	struct amdgpu_user_fence uf;
+};
 
-	struct amdgpu_ring *ring;
-	struct mutex job_lock;
-	struct work_struct job_work;
-	int (*prepare_job)(struct amdgpu_cs_parser *sched_job);
-	int (*run_job)(struct amdgpu_cs_parser *sched_job);
-	int (*free_job)(struct amdgpu_cs_parser *sched_job);
-	struct amd_sched_fence *s_fence;
+struct amdgpu_job {
+	struct amd_sched_job    base;
+	struct amdgpu_device	*adev;
+	struct amdgpu_ib	*ibs;
+	uint32_t		num_ibs;
+	struct mutex            job_lock;
+	struct amdgpu_user_fence uf;
+	int (*free_job)(struct amdgpu_job *sched_job);
 };
 
 static inline u32 amdgpu_get_ib_value(struct amdgpu_cs_parser *p, uint32_t ib_idx, int idx)
@@ -2241,7 +2247,7 @@ static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
 #define amdgpu_display_add_connector(adev, ci, sd, ct, ib, coi, h, r) (adev)->mode_info.funcs->add_connector((adev), (ci), (sd), (ct), (ib), (coi), (h), (r))
 #define amdgpu_display_stop_mc_access(adev, s) (adev)->mode_info.funcs->stop_mc_access((adev), (s))
 #define amdgpu_display_resume_mc_access(adev, s) (adev)->mode_info.funcs->resume_mc_access((adev), (s))
-#define amdgpu_emit_copy_buffer(adev, r, s, d, b) (adev)->mman.buffer_funcs->emit_copy_buffer((r), (s), (d), (b))
+#define amdgpu_emit_copy_buffer(adev, ib, s, d, b) (adev)->mman.buffer_funcs->emit_copy_buffer((ib),  (s), (d), (b))
 #define amdgpu_emit_fill_buffer(adev, r, s, d, b) (adev)->mman.buffer_funcs->emit_fill_buffer((r), (s), (d), (b))
 #define amdgpu_dpm_get_temperature(adev) (adev)->pm.funcs->get_temperature((adev))
 #define amdgpu_dpm_pre_set_power_state(adev) (adev)->pm.funcs->pre_set_power_state((adev))
@@ -2343,7 +2349,7 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
 		      struct amdgpu_sync *sync);
 void amdgpu_vm_flush(struct amdgpu_ring *ring,
 		     struct amdgpu_vm *vm,
-		     struct amdgpu_fence *updates);
+		     struct fence *updates);
 void amdgpu_vm_fence(struct amdgpu_device *adev,
 		     struct amdgpu_vm *vm,
 		     struct amdgpu_fence *fence);
@@ -2373,7 +2379,7 @@ int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
 		       uint64_t addr);
 void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
 		      struct amdgpu_bo_va *bo_va);
-
+int amdgpu_vm_free_job(struct amdgpu_job *job);
 /*
  * functions used by amdgpu_encoder.c
  */
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c
index 759482e4300d..98d59ee640ce 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c
@@ -33,7 +33,7 @@ static int amdgpu_benchmark_do_move(struct amdgpu_device *adev, unsigned size,
 {
 	unsigned long start_jiffies;
 	unsigned long end_jiffies;
-	struct amdgpu_fence *fence = NULL;
+	struct fence *fence = NULL;
 	int i, r;
 
 	start_jiffies = jiffies;
@@ -42,17 +42,17 @@ static int amdgpu_benchmark_do_move(struct amdgpu_device *adev, unsigned size,
 		r = amdgpu_copy_buffer(ring, saddr, daddr, size, NULL, &fence);
 		if (r)
 			goto exit_do_move;
-		r = fence_wait(&fence->base, false);
+		r = fence_wait(fence, false);
 		if (r)
 			goto exit_do_move;
-		amdgpu_fence_unref(&fence);
+		fence_put(fence);
 	}
 	end_jiffies = jiffies;
 	r = jiffies_to_msecs(end_jiffies - start_jiffies);
 
 exit_do_move:
 	if (fence)
-		amdgpu_fence_unref(&fence);
+		fence_put(fence);
 	return r;
 }
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c
index 7eed523bf28f..f82a2dd83874 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c
@@ -62,39 +62,6 @@ static int amdgpu_bo_list_create(struct amdgpu_fpriv *fpriv,
 	return 0;
 }
 
-struct amdgpu_bo_list *
-amdgpu_bo_list_clone(struct amdgpu_bo_list *list)
-{
-	struct amdgpu_bo_list *result;
-	unsigned i;
-
-	result = kmalloc(sizeof(struct amdgpu_bo_list), GFP_KERNEL);
-	if (!result)
-		return NULL;
-
-	result->array = drm_calloc_large(list->num_entries,
-		sizeof(struct amdgpu_bo_list_entry));
-	if (!result->array) {
-		kfree(result);
-		return NULL;
-	}
-
-	mutex_init(&result->lock);
-	result->gds_obj = list->gds_obj;
-	result->gws_obj = list->gws_obj;
-	result->oa_obj = list->oa_obj;
-	result->has_userptr = list->has_userptr;
-	result->num_entries = list->num_entries;
-
-	memcpy(result->array, list->array, list->num_entries *
-	       sizeof(struct amdgpu_bo_list_entry));
-
-	for (i = 0; i < result->num_entries; ++i)
-		amdgpu_bo_ref(result->array[i].robj);
-
-	return result;
-}
-
 static void amdgpu_bo_list_destroy(struct amdgpu_fpriv *fpriv, int id)
 {
 	struct amdgpu_bo_list *list;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_connectors.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_connectors.c
index 27df17a0e620..89c3dd62ba21 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_connectors.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_connectors.c
@@ -75,6 +75,11 @@ void amdgpu_connector_hotplug(struct drm_connector *connector)
 			if (!amdgpu_display_hpd_sense(adev, amdgpu_connector->hpd.hpd)) {
 				drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
 			} else if (amdgpu_atombios_dp_needs_link_train(amdgpu_connector)) {
+				/* Don't try to start link training before we
+				 * have the dpcd */
+				if (!amdgpu_atombios_dp_get_dpcd(amdgpu_connector))
+					return;
+
 				/* set it to OFF so that drm_helper_connector_dpms()
 				 * won't return immediately since the current state
 				 * is ON at this point.
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
index e4424b4db5d3..6a206f15635f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
@@ -126,19 +126,6 @@ int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
 	return 0;
 }
 
-static void amdgpu_job_work_func(struct work_struct *work)
-{
-	struct amdgpu_cs_parser *sched_job =
-		container_of(work, struct amdgpu_cs_parser,
-			     job_work);
-	mutex_lock(&sched_job->job_lock);
-	if (sched_job->free_job)
-		sched_job->free_job(sched_job);
-	mutex_unlock(&sched_job->job_lock);
-	/* after processing job, free memory */
-	fence_put(&sched_job->s_fence->base);
-	kfree(sched_job);
-}
 struct amdgpu_cs_parser *amdgpu_cs_parser_create(struct amdgpu_device *adev,
                                                struct drm_file *filp,
                                                struct amdgpu_ctx *ctx,
@@ -157,10 +144,6 @@ struct amdgpu_cs_parser *amdgpu_cs_parser_create(struct amdgpu_device *adev,
 	parser->ctx = ctx;
 	parser->ibs = ibs;
 	parser->num_ibs = num_ibs;
-	if (amdgpu_enable_scheduler) {
-		mutex_init(&parser->job_lock);
-		INIT_WORK(&parser->job_work, amdgpu_job_work_func);
-	}
 	for (i = 0; i < num_ibs; i++)
 		ibs[i].ctx = ctx;
 
@@ -173,7 +156,6 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 	uint64_t *chunk_array_user;
 	uint64_t *chunk_array = NULL;
 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
-	struct amdgpu_bo_list *bo_list = NULL;
 	unsigned size, i;
 	int r = 0;
 
@@ -185,20 +167,7 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 		r = -EINVAL;
 		goto out;
 	}
-	bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
-	if (!amdgpu_enable_scheduler)
-		p->bo_list = bo_list;
-	else {
-		if (bo_list && !bo_list->has_userptr) {
-			p->bo_list = amdgpu_bo_list_clone(bo_list);
-			amdgpu_bo_list_put(bo_list);
-			if (!p->bo_list)
-				return -ENOMEM;
-		} else if (bo_list && bo_list->has_userptr)
-			p->bo_list = bo_list;
-		else
-			p->bo_list = NULL;
-	}
+	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
 
 	/* get chunks */
 	INIT_LIST_HEAD(&p->validated);
@@ -291,7 +260,7 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 	}
 
 
-	p->ibs = kmalloc_array(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
+	p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
 	if (!p->ibs)
 		r = -ENOMEM;
 
@@ -498,25 +467,24 @@ static void amdgpu_cs_parser_fini_late(struct amdgpu_cs_parser *parser)
 	unsigned i;
 	if (parser->ctx)
 		amdgpu_ctx_put(parser->ctx);
-	if (parser->bo_list) {
-		if (amdgpu_enable_scheduler && !parser->bo_list->has_userptr)
-			amdgpu_bo_list_free(parser->bo_list);
-		else
-			amdgpu_bo_list_put(parser->bo_list);
-	}
+	if (parser->bo_list)
+		amdgpu_bo_list_put(parser->bo_list);
+
 	drm_free_large(parser->vm_bos);
 	for (i = 0; i < parser->nchunks; i++)
 		drm_free_large(parser->chunks[i].kdata);
 	kfree(parser->chunks);
-	if (parser->ibs)
-		for (i = 0; i < parser->num_ibs; i++)
-			amdgpu_ib_free(parser->adev, &parser->ibs[i]);
-	kfree(parser->ibs);
-	if (parser->uf.bo)
-		drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base);
-
 	if (!amdgpu_enable_scheduler)
-		kfree(parser);
+	{
+		if (parser->ibs)
+			for (i = 0; i < parser->num_ibs; i++)
+				amdgpu_ib_free(parser->adev, &parser->ibs[i]);
+		kfree(parser->ibs);
+		if (parser->uf.bo)
+			drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base);
+	}
+
+	kfree(parser);
 }
 
 /**
@@ -533,12 +501,6 @@ static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bo
        amdgpu_cs_parser_fini_late(parser);
 }
 
-static int amdgpu_cs_parser_free_job(struct amdgpu_cs_parser *sched_job)
-{
-       amdgpu_cs_parser_fini_late(sched_job);
-       return 0;
-}
-
 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
 				   struct amdgpu_vm *vm)
 {
@@ -810,68 +772,16 @@ static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
 	return 0;
 }
 
-static int amdgpu_cs_parser_prepare_job(struct amdgpu_cs_parser *sched_job)
+static int amdgpu_cs_free_job(struct amdgpu_job *sched_job)
 {
-	int r, i;
-	struct amdgpu_cs_parser *parser = sched_job;
-	struct amdgpu_device *adev = sched_job->adev;
-	bool reserved_buffers = false;
-
-	r = amdgpu_cs_parser_relocs(parser);
-	if (r) {
-		if (r != -ERESTARTSYS) {
-			if (r == -ENOMEM)
-				DRM_ERROR("Not enough memory for command submission!\n");
-			else
-				DRM_ERROR("Failed to process the buffer list %d!\n", r);
-		}
-	}
-
-	if (!r) {
-		reserved_buffers = true;
-		r = amdgpu_cs_ib_fill(adev, parser);
-	}
-	if (!r) {
-		r = amdgpu_cs_dependencies(adev, parser);
-		if (r)
-			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
-	}
-	if (r) {
-		amdgpu_cs_parser_fini(parser, r, reserved_buffers);
-		return r;
-	}
-
-	for (i = 0; i < parser->num_ibs; i++)
-		trace_amdgpu_cs(parser, i);
-
-	r = amdgpu_cs_ib_vm_chunk(adev, parser);
-	return r;
-}
-
-static struct amdgpu_ring *amdgpu_cs_parser_get_ring(
-	struct amdgpu_device *adev,
-	struct amdgpu_cs_parser *parser)
-{
-	int i, r;
-
-	struct amdgpu_cs_chunk *chunk;
-	struct drm_amdgpu_cs_chunk_ib *chunk_ib;
-	struct amdgpu_ring *ring;
-	for (i = 0; i < parser->nchunks; i++) {
-		chunk = &parser->chunks[i];
-		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
-
-		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
-			continue;
-
-		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
-				       chunk_ib->ip_instance, chunk_ib->ring,
-				       &ring);
-		if (r)
-			return NULL;
-		break;
-	}
-	return ring;
+	int i;
+	if (sched_job->ibs)
+		for (i = 0; i < sched_job->num_ibs; i++)
+			amdgpu_ib_free(sched_job->adev, &sched_job->ibs[i]);
+	kfree(sched_job->ibs);
+	if (sched_job->uf.bo)
+		drm_gem_object_unreference_unlocked(&sched_job->uf.bo->gem_base);
+	return 0;
 }
 
 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
@@ -879,7 +789,8 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	struct amdgpu_device *adev = dev->dev_private;
 	union drm_amdgpu_cs *cs = data;
 	struct amdgpu_cs_parser *parser;
-	int r;
+	bool reserved_buffers = false;
+	int i, r;
 
 	down_read(&adev->exclusive_lock);
 	if (!adev->accel_working) {
@@ -899,44 +810,79 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 		return r;
 	}
 
-	if (amdgpu_enable_scheduler && parser->num_ibs) {
-		struct amdgpu_ring * ring =
-			amdgpu_cs_parser_get_ring(adev, parser);
-		r = amdgpu_cs_parser_prepare_job(parser);
+	r = amdgpu_cs_parser_relocs(parser);
+	if (r == -ENOMEM)
+		DRM_ERROR("Not enough memory for command submission!\n");
+	else if (r && r != -ERESTARTSYS)
+		DRM_ERROR("Failed to process the buffer list %d!\n", r);
+	else if (!r) {
+		reserved_buffers = true;
+		r = amdgpu_cs_ib_fill(adev, parser);
+	}
+
+	if (!r) {
+		r = amdgpu_cs_dependencies(adev, parser);
 		if (r)
-			goto out;
-		parser->ring = ring;
-		parser->free_job = amdgpu_cs_parser_free_job;
-		mutex_lock(&parser->job_lock);
-		r = amd_sched_push_job(ring->scheduler,
-				       &parser->ctx->rings[ring->idx].entity,
-				       parser,
-				       &parser->s_fence);
+			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
+	}
+
+	if (r)
+		goto out;
+
+	for (i = 0; i < parser->num_ibs; i++)
+		trace_amdgpu_cs(parser, i);
+
+	r = amdgpu_cs_ib_vm_chunk(adev, parser);
+	if (r)
+		goto out;
+
+	if (amdgpu_enable_scheduler && parser->num_ibs) {
+		struct amdgpu_job *job;
+		struct amdgpu_ring * ring =  parser->ibs->ring;
+		job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL);
+		if (!job)
+			return -ENOMEM;
+		job->base.sched = ring->scheduler;
+		job->base.s_entity = &parser->ctx->rings[ring->idx].entity;
+		job->adev = parser->adev;
+		job->ibs = parser->ibs;
+		job->num_ibs = parser->num_ibs;
+		job->base.owner = parser->filp;
+		mutex_init(&job->job_lock);
+		if (job->ibs[job->num_ibs - 1].user) {
+			memcpy(&job->uf,  &parser->uf,
+			       sizeof(struct amdgpu_user_fence));
+			job->ibs[job->num_ibs - 1].user = &job->uf;
+		}
+
+		job->free_job = amdgpu_cs_free_job;
+		mutex_lock(&job->job_lock);
+		r = amd_sched_entity_push_job((struct amd_sched_job *)job);
 		if (r) {
-			mutex_unlock(&parser->job_lock);
+			mutex_unlock(&job->job_lock);
+			amdgpu_cs_free_job(job);
+			kfree(job);
 			goto out;
 		}
-		parser->ibs[parser->num_ibs - 1].sequence =
+		cs->out.handle =
 			amdgpu_ctx_add_fence(parser->ctx, ring,
-					     &parser->s_fence->base,
-					     parser->s_fence->v_seq);
-		cs->out.handle = parser->s_fence->v_seq;
+					     &job->base.s_fence->base);
+		parser->ibs[parser->num_ibs - 1].sequence = cs->out.handle;
+
 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
 		ttm_eu_fence_buffer_objects(&parser->ticket,
 				&parser->validated,
-				&parser->s_fence->base);
+				&job->base.s_fence->base);
 
-		mutex_unlock(&parser->job_lock);
+		mutex_unlock(&job->job_lock);
+		amdgpu_cs_parser_fini_late(parser);
 		up_read(&adev->exclusive_lock);
 		return 0;
 	}
-	r = amdgpu_cs_parser_prepare_job(parser);
-	if (r)
-		goto out;
 
 	cs->out.handle = parser->ibs[parser->num_ibs - 1].sequence;
 out:
-	amdgpu_cs_parser_fini(parser, r, true);
+	amdgpu_cs_parser_fini(parser, r, reserved_buffers);
 	up_read(&adev->exclusive_lock);
 	r = amdgpu_cs_handle_lockup(adev, r);
 	return r;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
index 08bc7722ddb8..20cbc4eb5a6f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
@@ -229,17 +229,13 @@ int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
 }
 
 uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
-			      struct fence *fence, uint64_t queued_seq)
+			      struct fence *fence)
 {
 	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
-	uint64_t seq = 0;
+	uint64_t seq = cring->sequence;
 	unsigned idx = 0;
 	struct fence *other = NULL;
 
-	if (amdgpu_enable_scheduler)
-		seq = queued_seq;
-	else
-		seq = cring->sequence;
 	idx = seq % AMDGPU_CTX_MAX_CS_PENDING;
 	other = cring->fences[idx];
 	if (other) {
@@ -253,8 +249,7 @@ uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
 
 	spin_lock(&ctx->ring_lock);
 	cring->fences[idx] = fence;
-	if (!amdgpu_enable_scheduler)
-		cring->sequence++;
+	cring->sequence++;
 	spin_unlock(&ctx->ring_lock);
 
 	fence_put(other);
@@ -267,21 +262,16 @@ struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 {
 	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
 	struct fence *fence;
-	uint64_t queued_seq;
 
 	spin_lock(&ctx->ring_lock);
-	if (amdgpu_enable_scheduler)
-		queued_seq = amd_sched_next_queued_seq(&cring->entity);
-	else
-		queued_seq = cring->sequence;
 
-	if (seq >= queued_seq) {
+	if (seq >= cring->sequence) {
 		spin_unlock(&ctx->ring_lock);
 		return ERR_PTR(-EINVAL);
 	}
 
 
-	if (seq + AMDGPU_CTX_MAX_CS_PENDING < queued_seq) {
+	if (seq + AMDGPU_CTX_MAX_CS_PENDING < cring->sequence) {
 		spin_unlock(&ctx->ring_lock);
 		return NULL;
 	}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index e6fa27805207..0fcc0bd1622c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -49,9 +49,10 @@
 /*
  * KMS wrapper.
  * - 3.0.0 - initial driver
+ * - 3.1.0 - allow reading more status registers (GRBM, SRBM, SDMA, CP)
  */
 #define KMS_DRIVER_MAJOR	3
-#define KMS_DRIVER_MINOR	0
+#define KMS_DRIVER_MINOR	1
 #define KMS_DRIVER_PATCHLEVEL	0
 
 int amdgpu_vram_limit = 0;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index 98500f1756f7..f446bf2fedc9 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -626,10 +626,10 @@ void amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
 	ring->fence_drv.ring = ring;
 
 	if (amdgpu_enable_scheduler) {
-		ring->scheduler = amd_sched_create((void *)ring->adev,
-						   &amdgpu_sched_ops,
-						   ring->idx, 5, 0,
-						   amdgpu_sched_hw_submission);
+		ring->scheduler = amd_sched_create(&amdgpu_sched_ops,
+						   ring->idx,
+						   amdgpu_sched_hw_submission,
+						   (void *)ring->adev);
 		if (!ring->scheduler)
 			DRM_ERROR("Failed to create scheduler on ring %d.\n",
 				  ring->idx);
@@ -836,22 +836,37 @@ static inline bool amdgpu_test_signaled(struct amdgpu_fence *fence)
 	return test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags);
 }
 
-static inline bool amdgpu_test_signaled_any(struct amdgpu_fence **fences)
+static bool amdgpu_test_signaled_any(struct fence **fences, uint32_t count)
 {
 	int idx;
-	struct amdgpu_fence *fence;
+	struct fence *fence;
 
-	idx = 0;
-	for (idx = 0; idx < AMDGPU_MAX_RINGS; ++idx) {
+	for (idx = 0; idx < count; ++idx) {
 		fence = fences[idx];
 		if (fence) {
-			if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags))
+			if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->flags))
 				return true;
 		}
 	}
 	return false;
 }
 
+static bool amdgpu_test_signaled_all(struct fence **fences, uint32_t count)
+{
+	int idx;
+	struct fence *fence;
+
+	for (idx = 0; idx < count; ++idx) {
+		fence = fences[idx];
+		if (fence) {
+			if (!test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->flags))
+				return false;
+		}
+	}
+
+	return true;
+}
+
 struct amdgpu_wait_cb {
 	struct fence_cb base;
 	struct task_struct *task;
@@ -867,33 +882,56 @@ static void amdgpu_fence_wait_cb(struct fence *fence, struct fence_cb *cb)
 static signed long amdgpu_fence_default_wait(struct fence *f, bool intr,
 					     signed long t)
 {
-	struct amdgpu_fence *array[AMDGPU_MAX_RINGS];
 	struct amdgpu_fence *fence = to_amdgpu_fence(f);
 	struct amdgpu_device *adev = fence->ring->adev;
 
-	memset(&array[0], 0, sizeof(array));
-	array[0] = fence;
-
-	return amdgpu_fence_wait_any(adev, array, intr, t);
+	return amdgpu_fence_wait_multiple(adev, &f, 1, false, intr, t);
 }
 
-/* wait until any fence in array signaled */
-signed long amdgpu_fence_wait_any(struct amdgpu_device *adev,
-				struct amdgpu_fence **array, bool intr, signed long t)
+/**
+ * Wait the fence array with timeout
+ *
+ * @adev:     amdgpu device
+ * @array:    the fence array with amdgpu fence pointer
+ * @count:    the number of the fence array
+ * @wait_all: the flag of wait all(true) or wait any(false)
+ * @intr:     when sleep, set the current task interruptable or not
+ * @t:        timeout to wait
+ *
+ * If wait_all is true, it will return when all fences are signaled or timeout.
+ * If wait_all is false, it will return when any fence is signaled or timeout.
+ */
+signed long amdgpu_fence_wait_multiple(struct amdgpu_device *adev,
+				       struct fence **array,
+				       uint32_t count,
+				       bool wait_all,
+				       bool intr,
+				       signed long t)
 {
 	long idx = 0;
-	struct amdgpu_wait_cb cb[AMDGPU_MAX_RINGS];
-	struct amdgpu_fence *fence;
+	struct amdgpu_wait_cb *cb;
+	struct fence *fence;
 
 	BUG_ON(!array);
 
-	for (idx = 0; idx < AMDGPU_MAX_RINGS; ++idx) {
+	cb = kcalloc(count, sizeof(struct amdgpu_wait_cb), GFP_KERNEL);
+	if (cb == NULL) {
+		t = -ENOMEM;
+		goto err_free_cb;
+	}
+
+	for (idx = 0; idx < count; ++idx) {
 		fence = array[idx];
 		if (fence) {
 			cb[idx].task = current;
-			if (fence_add_callback(&fence->base,
-					&cb[idx].base, amdgpu_fence_wait_cb))
-				return t; /* return if fence is already signaled */
+			if (fence_add_callback(fence,
+					&cb[idx].base, amdgpu_fence_wait_cb)) {
+				/* The fence is already signaled */
+				if (wait_all)
+					continue;
+				else
+					goto fence_rm_cb;
+			}
 		}
 	}
 
@@ -907,7 +945,9 @@ signed long amdgpu_fence_wait_any(struct amdgpu_device *adev,
 		 * amdgpu_test_signaled_any must be called after
 		 * set_current_state to prevent a race with wake_up_process
 		 */
-		if (amdgpu_test_signaled_any(array))
+		if (!wait_all && amdgpu_test_signaled_any(array, count))
+			break;
+		if (wait_all && amdgpu_test_signaled_all(array, count))
 			break;
 
 		if (adev->needs_reset) {
@@ -923,13 +963,16 @@ signed long amdgpu_fence_wait_any(struct amdgpu_device *adev,
 
 	__set_current_state(TASK_RUNNING);
 
-	idx = 0;
-	for (idx = 0; idx < AMDGPU_MAX_RINGS; ++idx) {
+fence_rm_cb:
+	for (idx = 0; idx < count; ++idx) {
 		fence = array[idx];
-		if (fence)
-			fence_remove_callback(&fence->base, &cb[idx].base);
+		if (fence && cb[idx].base.func)
+			fence_remove_callback(fence, &cb[idx].base);
 	}
 
+err_free_cb:
+	kfree(cb);
+
 	return t;
 }
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
index 5104e64e9ad8..c439735ee670 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
@@ -73,29 +73,12 @@ int amdgpu_ib_get(struct amdgpu_ring *ring, struct amdgpu_vm *vm,
 
 		if (!vm)
 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
-		else
-			ib->gpu_addr = 0;
-
-	} else {
-		ib->sa_bo = NULL;
-		ib->ptr = NULL;
-		ib->gpu_addr = 0;
 	}
 
 	amdgpu_sync_create(&ib->sync);
 
 	ib->ring = ring;
-	ib->fence = NULL;
-	ib->user = NULL;
 	ib->vm = vm;
-	ib->ctx = NULL;
-	ib->gds_base = 0;
-	ib->gds_size = 0;
-	ib->gws_base = 0;
-	ib->gws_size = 0;
-	ib->oa_base = 0;
-	ib->oa_size = 0;
-	ib->flags = 0;
 
 	return 0;
 }
@@ -110,8 +93,8 @@ int amdgpu_ib_get(struct amdgpu_ring *ring, struct amdgpu_vm *vm,
  */
 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib)
 {
-	amdgpu_sync_free(adev, &ib->sync, ib->fence);
-	amdgpu_sa_bo_free(adev, &ib->sa_bo, ib->fence);
+	amdgpu_sync_free(adev, &ib->sync, &ib->fence->base);
+	amdgpu_sa_bo_free(adev, &ib->sa_bo, &ib->fence->base);
 	amdgpu_fence_unref(&ib->fence);
 }
 
@@ -143,7 +126,6 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 	struct amdgpu_ring *ring;
 	struct amdgpu_ctx *ctx, *old_ctx;
 	struct amdgpu_vm *vm;
-	uint64_t sequence;
 	unsigned i;
 	int r = 0;
 
@@ -158,7 +140,11 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 		dev_err(adev->dev, "couldn't schedule ib\n");
 		return -EINVAL;
 	}
-
+	r = amdgpu_sync_wait(&ibs->sync);
+	if (r) {
+		dev_err(adev->dev, "IB sync failed (%d).\n", r);
+		return r;
+	}
 	r = amdgpu_ring_lock(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs);
 	if (r) {
 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
@@ -216,12 +202,9 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 		return r;
 	}
 
-	sequence = amdgpu_enable_scheduler ? ib->sequence : 0;
-
 	if (!amdgpu_enable_scheduler && ib->ctx)
 		ib->sequence = amdgpu_ctx_add_fence(ib->ctx, ring,
-						    &ib->fence->base,
-						    sequence);
+						    &ib->fence->base);
 
 	/* wrap the last IB with fence */
 	if (ib->user) {
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c
index 90044b254404..5c8a803acedc 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c
@@ -98,18 +98,12 @@ int amdgpu_ih_ring_init(struct amdgpu_device *adev, unsigned ring_size,
 			/* add 8 bytes for the rptr/wptr shadows and
 			 * add them to the end of the ring allocation.
 			 */
-			adev->irq.ih.ring = kzalloc(adev->irq.ih.ring_size + 8, GFP_KERNEL);
+			adev->irq.ih.ring = pci_alloc_consistent(adev->pdev,
+								 adev->irq.ih.ring_size + 8,
+								 &adev->irq.ih.rb_dma_addr);
 			if (adev->irq.ih.ring == NULL)
 				return -ENOMEM;
-			adev->irq.ih.rb_dma_addr = pci_map_single(adev->pdev,
-								  (void *)adev->irq.ih.ring,
-								  adev->irq.ih.ring_size,
-								  PCI_DMA_BIDIRECTIONAL);
-			if (pci_dma_mapping_error(adev->pdev, adev->irq.ih.rb_dma_addr)) {
-				dev_err(&adev->pdev->dev, "Failed to DMA MAP the IH RB page\n");
-				kfree((void *)adev->irq.ih.ring);
-				return -ENOMEM;
-			}
+			memset((void *)adev->irq.ih.ring, 0, adev->irq.ih.ring_size + 8);
 			adev->irq.ih.wptr_offs = (adev->irq.ih.ring_size / 4) + 0;
 			adev->irq.ih.rptr_offs = (adev->irq.ih.ring_size / 4) + 1;
 		}
@@ -149,9 +143,9 @@ void amdgpu_ih_ring_fini(struct amdgpu_device *adev)
 			/* add 8 bytes for the rptr/wptr shadows and
 			 * add them to the end of the ring allocation.
 			 */
-			pci_unmap_single(adev->pdev, adev->irq.ih.rb_dma_addr,
-					 adev->irq.ih.ring_size + 8, PCI_DMA_BIDIRECTIONAL);
-			kfree((void *)adev->irq.ih.ring);
+			pci_free_consistent(adev->pdev, adev->irq.ih.ring_size + 8,
+					    (void *)adev->irq.ih.ring,
+					    adev->irq.ih.rb_dma_addr);
 			adev->irq.ih.ring = NULL;
 		}
 	} else {
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
index 87da6b1848fd..22367939ebf1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
@@ -560,6 +560,8 @@ void amdgpu_driver_postclose_kms(struct drm_device *dev,
 	if (!fpriv)
 		return;
 
+	amdgpu_ctx_mgr_fini(&fpriv->ctx_mgr);
+
 	amdgpu_vm_fini(adev, &fpriv->vm);
 
 	idr_for_each_entry(&fpriv->bo_list_handles, list, handle)
@@ -568,8 +570,6 @@ void amdgpu_driver_postclose_kms(struct drm_device *dev,
 	idr_destroy(&fpriv->bo_list_handles);
 	mutex_destroy(&fpriv->bo_list_lock);
 
-	amdgpu_ctx_mgr_fini(&fpriv->ctx_mgr);
-
 	kfree(fpriv);
 	file_priv->driver_priv = NULL;
 }
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
index 238465a9ac55..6ea18dcec561 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
@@ -193,7 +193,7 @@ int amdgpu_sa_bo_new(struct amdgpu_device *adev,
 			    unsigned size, unsigned align);
 void amdgpu_sa_bo_free(struct amdgpu_device *adev,
 			      struct amdgpu_sa_bo **sa_bo,
-			      struct amdgpu_fence *fence);
+			      struct fence *fence);
 #if defined(CONFIG_DEBUG_FS)
 void amdgpu_sa_bo_dump_debug_info(struct amdgpu_sa_manager *sa_manager,
 					 struct seq_file *m);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c
index d6398cf45f24..b92525329d6c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c
@@ -139,6 +139,20 @@ int amdgpu_sa_bo_manager_suspend(struct amdgpu_device *adev,
 	return r;
 }
 
+static uint32_t amdgpu_sa_get_ring_from_fence(struct fence *f)
+{
+	struct amdgpu_fence *a_fence;
+	struct amd_sched_fence *s_fence;
+
+	s_fence = to_amd_sched_fence(f);
+	if (s_fence)
+		return s_fence->scheduler->ring_id;
+	a_fence = to_amdgpu_fence(f);
+	if (a_fence)
+		return a_fence->ring->idx;
+	return 0;
+}
+
 static void amdgpu_sa_bo_remove_locked(struct amdgpu_sa_bo *sa_bo)
 {
 	struct amdgpu_sa_manager *sa_manager = sa_bo->manager;
@@ -147,7 +161,7 @@ static void amdgpu_sa_bo_remove_locked(struct amdgpu_sa_bo *sa_bo)
 	}
 	list_del_init(&sa_bo->olist);
 	list_del_init(&sa_bo->flist);
-	amdgpu_fence_unref(&sa_bo->fence);
+	fence_put(sa_bo->fence);
 	kfree(sa_bo);
 }
 
@@ -161,7 +175,7 @@ static void amdgpu_sa_bo_try_free(struct amdgpu_sa_manager *sa_manager)
 	sa_bo = list_entry(sa_manager->hole->next, struct amdgpu_sa_bo, olist);
 	list_for_each_entry_safe_from(sa_bo, tmp, &sa_manager->olist, olist) {
 		if (sa_bo->fence == NULL ||
-		    !fence_is_signaled(&sa_bo->fence->base)) {
+		    !fence_is_signaled(sa_bo->fence)) {
 			return;
 		}
 		amdgpu_sa_bo_remove_locked(sa_bo);
@@ -246,7 +260,7 @@ static bool amdgpu_sa_event(struct amdgpu_sa_manager *sa_manager,
 }
 
 static bool amdgpu_sa_bo_next_hole(struct amdgpu_sa_manager *sa_manager,
-				   struct amdgpu_fence **fences,
+				   struct fence **fences,
 				   unsigned *tries)
 {
 	struct amdgpu_sa_bo *best_bo = NULL;
@@ -275,7 +289,7 @@ static bool amdgpu_sa_bo_next_hole(struct amdgpu_sa_manager *sa_manager,
 		sa_bo = list_first_entry(&sa_manager->flist[i],
 					 struct amdgpu_sa_bo, flist);
 
-		if (!fence_is_signaled(&sa_bo->fence->base)) {
+		if (!fence_is_signaled(sa_bo->fence)) {
 			fences[i] = sa_bo->fence;
 			continue;
 		}
@@ -299,7 +313,8 @@ static bool amdgpu_sa_bo_next_hole(struct amdgpu_sa_manager *sa_manager,
 	}
 
 	if (best_bo) {
-		++tries[best_bo->fence->ring->idx];
+		uint32_t idx = amdgpu_sa_get_ring_from_fence(best_bo->fence);
+		++tries[idx];
 		sa_manager->hole = best_bo->olist.prev;
 
 		/* we knew that this one is signaled,
@@ -315,7 +330,7 @@ int amdgpu_sa_bo_new(struct amdgpu_device *adev,
 		     struct amdgpu_sa_bo **sa_bo,
 		     unsigned size, unsigned align)
 {
-	struct amdgpu_fence *fences[AMDGPU_MAX_RINGS];
+	struct fence *fences[AMDGPU_MAX_RINGS];
 	unsigned tries[AMDGPU_MAX_RINGS];
 	int i, r;
 	signed long t;
@@ -352,7 +367,8 @@ int amdgpu_sa_bo_new(struct amdgpu_device *adev,
 		} while (amdgpu_sa_bo_next_hole(sa_manager, fences, tries));
 
 		spin_unlock(&sa_manager->wq.lock);
-		t = amdgpu_fence_wait_any(adev, fences, false, MAX_SCHEDULE_TIMEOUT);
+		t = amdgpu_fence_wait_multiple(adev, fences, AMDGPU_MAX_RINGS, false, false,
+						MAX_SCHEDULE_TIMEOUT);
 		r = (t > 0) ? 0 : t;
 		spin_lock(&sa_manager->wq.lock);
 		/* if we have nothing to wait for block */
@@ -372,7 +388,7 @@ int amdgpu_sa_bo_new(struct amdgpu_device *adev,
 }
 
 void amdgpu_sa_bo_free(struct amdgpu_device *adev, struct amdgpu_sa_bo **sa_bo,
-		       struct amdgpu_fence *fence)
+		       struct fence *fence)
 {
 	struct amdgpu_sa_manager *sa_manager;
 
@@ -382,10 +398,11 @@ void amdgpu_sa_bo_free(struct amdgpu_device *adev, struct amdgpu_sa_bo **sa_bo,
 
 	sa_manager = (*sa_bo)->manager;
 	spin_lock(&sa_manager->wq.lock);
-	if (fence && !fence_is_signaled(&fence->base)) {
-		(*sa_bo)->fence = amdgpu_fence_ref(fence);
-		list_add_tail(&(*sa_bo)->flist,
-			      &sa_manager->flist[fence->ring->idx]);
+	if (fence && !fence_is_signaled(fence)) {
+		uint32_t idx;
+		(*sa_bo)->fence = fence_get(fence);
+		idx = amdgpu_sa_get_ring_from_fence(fence);
+		list_add_tail(&(*sa_bo)->flist, &sa_manager->flist[idx]);
 	} else {
 		amdgpu_sa_bo_remove_locked(*sa_bo);
 	}
@@ -412,8 +429,16 @@ void amdgpu_sa_bo_dump_debug_info(struct amdgpu_sa_manager *sa_manager,
 		seq_printf(m, "[0x%010llx 0x%010llx] size %8lld",
 			   soffset, eoffset, eoffset - soffset);
 		if (i->fence) {
-			seq_printf(m, " protected by 0x%016llx on ring %d",
-				   i->fence->seq, i->fence->ring->idx);
+			struct amdgpu_fence *a_fence = to_amdgpu_fence(i->fence);
+			struct amd_sched_fence *s_fence = to_amd_sched_fence(i->fence);
+			if (a_fence)
+				seq_printf(m, " protected by 0x%016llx on ring %d",
+					   a_fence->seq, a_fence->ring->idx);
+			if (s_fence)
+				seq_printf(m, " protected by 0x%016x on ring %d",
+					   s_fence->base.seqno,
+					   s_fence->scheduler->ring_id);
+
 		}
 		seq_printf(m, "\n");
 	}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c
index a86e38158afa..f93fb3541488 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c
@@ -27,55 +27,28 @@
 #include <drm/drmP.h>
 #include "amdgpu.h"
 
-static int amdgpu_sched_prepare_job(struct amd_gpu_scheduler *sched,
-				    struct amd_sched_entity *entity,
-				    struct amd_sched_job *job)
+static struct fence *amdgpu_sched_run_job(struct amd_sched_job *job)
 {
-	int r = 0;
-	struct amdgpu_cs_parser *sched_job;
-	if (!job || !job->data) {
-		DRM_ERROR("job is null\n");
-		return -EINVAL;
-	}
-
-	sched_job = (struct amdgpu_cs_parser *)job->data;
-	if (sched_job->prepare_job) {
-		r = sched_job->prepare_job(sched_job);
-		if (r) {
-			DRM_ERROR("Prepare job error\n");
-			schedule_work(&sched_job->job_work);
-		}
-	}
-	return r;
-}
-
-static struct fence *amdgpu_sched_run_job(struct amd_gpu_scheduler *sched,
-					  struct amd_sched_entity *entity,
-					  struct amd_sched_job *job)
-{
-	int r = 0;
-	struct amdgpu_cs_parser *sched_job;
+	struct amdgpu_job *sched_job;
 	struct amdgpu_fence *fence;
+	int r;
 
-	if (!job || !job->data) {
+	if (!job) {
 		DRM_ERROR("job is null\n");
 		return NULL;
 	}
-	sched_job = (struct amdgpu_cs_parser *)job->data;
+	sched_job = (struct amdgpu_job *)job;
 	mutex_lock(&sched_job->job_lock);
 	r = amdgpu_ib_schedule(sched_job->adev,
 			       sched_job->num_ibs,
 			       sched_job->ibs,
-			       sched_job->filp);
+			       sched_job->base.owner);
 	if (r)
 		goto err;
 	fence = amdgpu_fence_ref(sched_job->ibs[sched_job->num_ibs - 1].fence);
 
-	if (sched_job->run_job) {
-		r = sched_job->run_job(sched_job);
-		if (r)
-			goto err;
-	}
+	if (sched_job->free_job)
+		sched_job->free_job(sched_job);
 
 	mutex_unlock(&sched_job->job_lock);
 	return &fence->base;
@@ -83,25 +56,25 @@ static struct fence *amdgpu_sched_run_job(struct amd_gpu_scheduler *sched,
 err:
 	DRM_ERROR("Run job error\n");
 	mutex_unlock(&sched_job->job_lock);
-	schedule_work(&sched_job->job_work);
+	job->sched->ops->process_job(job);
 	return NULL;
 }
 
-static void amdgpu_sched_process_job(struct amd_gpu_scheduler *sched,
-				     struct amd_sched_job *job)
+static void amdgpu_sched_process_job(struct amd_sched_job *job)
 {
-	struct amdgpu_cs_parser *sched_job;
+	struct amdgpu_job *sched_job;
 
-	if (!job || !job->data) {
+	if (!job) {
 		DRM_ERROR("job is null\n");
 		return;
 	}
-	sched_job = (struct amdgpu_cs_parser *)job->data;
-	schedule_work(&sched_job->job_work);
+	sched_job = (struct amdgpu_job *)job;
+	/* after processing job, free memory */
+	fence_put(&sched_job->base.s_fence->base);
+	kfree(sched_job);
 }
 
 struct amd_sched_backend_ops amdgpu_sched_ops = {
-	.prepare_job = amdgpu_sched_prepare_job,
 	.run_job = amdgpu_sched_run_job,
 	.process_job = amdgpu_sched_process_job
 };
@@ -110,36 +83,39 @@ int amdgpu_sched_ib_submit_kernel_helper(struct amdgpu_device *adev,
 					 struct amdgpu_ring *ring,
 					 struct amdgpu_ib *ibs,
 					 unsigned num_ibs,
-					 int (*free_job)(struct amdgpu_cs_parser *),
+					 int (*free_job)(struct amdgpu_job *),
 					 void *owner,
 					 struct fence **f)
 {
 	int r = 0;
 	if (amdgpu_enable_scheduler) {
-		struct amdgpu_cs_parser *sched_job =
-			amdgpu_cs_parser_create(adev, owner, &adev->kernel_ctx,
-						ibs, num_ibs);
-		if(!sched_job) {
+		struct amdgpu_job *job =
+			kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL);
+		if (!job)
 			return -ENOMEM;
-		}
-		sched_job->free_job = free_job;
-		mutex_lock(&sched_job->job_lock);
-		r = amd_sched_push_job(ring->scheduler,
-				       &adev->kernel_ctx.rings[ring->idx].entity,
-				       sched_job, &sched_job->s_fence);
+		job->base.sched = ring->scheduler;
+		job->base.s_entity = &adev->kernel_ctx.rings[ring->idx].entity;
+		job->adev = adev;
+		job->ibs = ibs;
+		job->num_ibs = num_ibs;
+		job->base.owner = owner;
+		mutex_init(&job->job_lock);
+		job->free_job = free_job;
+		mutex_lock(&job->job_lock);
+		r = amd_sched_entity_push_job((struct amd_sched_job *)job);
 		if (r) {
-			mutex_unlock(&sched_job->job_lock);
-			kfree(sched_job);
+			mutex_unlock(&job->job_lock);
+			kfree(job);
 			return r;
 		}
-		ibs[num_ibs - 1].sequence = sched_job->s_fence->v_seq;
-		*f = fence_get(&sched_job->s_fence->base);
-		mutex_unlock(&sched_job->job_lock);
+		*f = fence_get(&job->base.s_fence->base);
+		mutex_unlock(&job->job_lock);
 	} else {
 		r = amdgpu_ib_schedule(adev, num_ibs, ibs, owner);
 		if (r)
 			return r;
 		*f = fence_get(&ibs[num_ibs - 1].fence->base);
 	}
+
 	return 0;
 }
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_semaphore.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_semaphore.c
index d6d41a42ab65..ff3ca52ec6fe 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_semaphore.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_semaphore.c
@@ -87,7 +87,7 @@ bool amdgpu_semaphore_emit_wait(struct amdgpu_ring *ring,
 
 void amdgpu_semaphore_free(struct amdgpu_device *adev,
 			   struct amdgpu_semaphore **semaphore,
-			   struct amdgpu_fence *fence)
+			   struct fence *fence)
 {
 	if (semaphore == NULL || *semaphore == NULL) {
 		return;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
index 7cb711fc1ee2..4fffb2539331 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
@@ -32,6 +32,11 @@
 #include "amdgpu.h"
 #include "amdgpu_trace.h"
 
+struct amdgpu_sync_entry {
+	struct hlist_node	node;
+	struct fence		*fence;
+};
+
 /**
  * amdgpu_sync_create - zero init sync object
  *
@@ -49,9 +54,33 @@ void amdgpu_sync_create(struct amdgpu_sync *sync)
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
 		sync->sync_to[i] = NULL;
 
+	hash_init(sync->fences);
 	sync->last_vm_update = NULL;
 }
 
+static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f)
+{
+	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
+	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+
+	if (a_fence)
+		return a_fence->ring->adev == adev;
+	if (s_fence)
+		return (struct amdgpu_device *)s_fence->scheduler->priv == adev;
+	return false;
+}
+
+static bool amdgpu_sync_test_owner(struct fence *f, void *owner)
+{
+	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
+	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+	if (s_fence)
+		return s_fence->owner == owner;
+	if (a_fence)
+		return a_fence->owner == owner;
+	return false;
+}
+
 /**
  * amdgpu_sync_fence - remember to sync to this fence
  *
@@ -62,28 +91,54 @@ void amdgpu_sync_create(struct amdgpu_sync *sync)
 int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 		      struct fence *f)
 {
+	struct amdgpu_sync_entry *e;
 	struct amdgpu_fence *fence;
 	struct amdgpu_fence *other;
+	struct fence *tmp, *later;
 
 	if (!f)
 		return 0;
 
+	if (amdgpu_sync_same_dev(adev, f) &&
+	    amdgpu_sync_test_owner(f, AMDGPU_FENCE_OWNER_VM)) {
+		if (sync->last_vm_update) {
+			tmp = sync->last_vm_update;
+			BUG_ON(f->context != tmp->context);
+			later = (f->seqno - tmp->seqno <= INT_MAX) ? f : tmp;
+			sync->last_vm_update = fence_get(later);
+			fence_put(tmp);
+		} else
+			sync->last_vm_update = fence_get(f);
+	}
+
 	fence = to_amdgpu_fence(f);
-	if (!fence || fence->ring->adev != adev)
-		return fence_wait(f, true);
+	if (!fence || fence->ring->adev != adev) {
+		hash_for_each_possible(sync->fences, e, node, f->context) {
+			struct fence *new;
+			if (unlikely(e->fence->context != f->context))
+				continue;
+			new = fence_get(fence_later(e->fence, f));
+			if (new) {
+				fence_put(e->fence);
+				e->fence = new;
+			}
+			return 0;
+		}
+
+		e = kmalloc(sizeof(struct amdgpu_sync_entry), GFP_KERNEL);
+		if (!e)
+			return -ENOMEM;
+
+		hash_add(sync->fences, &e->node, f->context);
+		e->fence = fence_get(f);
+		return 0;
+	}
 
 	other = sync->sync_to[fence->ring->idx];
 	sync->sync_to[fence->ring->idx] = amdgpu_fence_ref(
 		amdgpu_fence_later(fence, other));
 	amdgpu_fence_unref(&other);
 
-	if (fence->owner == AMDGPU_FENCE_OWNER_VM) {
-		other = sync->last_vm_update;
-		sync->last_vm_update = amdgpu_fence_ref(
-			amdgpu_fence_later(fence, other));
-		amdgpu_fence_unref(&other);
-	}
-
 	return 0;
 }
 
@@ -147,6 +202,24 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
 	return r;
 }
 
+int amdgpu_sync_wait(struct amdgpu_sync *sync)
+{
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
+	int i, r;
+
+	hash_for_each_safe(sync->fences, i, tmp, e, node) {
+		r = fence_wait(e->fence, false);
+		if (r)
+			return r;
+
+		hash_del(&e->node);
+		fence_put(e->fence);
+		kfree(e);
+	}
+	return 0;
+}
+
 /**
  * amdgpu_sync_rings - sync ring to all registered fences
  *
@@ -234,15 +307,23 @@ int amdgpu_sync_rings(struct amdgpu_sync *sync,
  */
 void amdgpu_sync_free(struct amdgpu_device *adev,
 		      struct amdgpu_sync *sync,
-		      struct amdgpu_fence *fence)
+		      struct fence *fence)
 {
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
 	unsigned i;
 
+	hash_for_each_safe(sync->fences, i, tmp, e, node) {
+		hash_del(&e->node);
+		fence_put(e->fence);
+		kfree(e);
+	}
+
 	for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
 		amdgpu_semaphore_free(adev, &sync->semaphores[i], fence);
 
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
 		amdgpu_fence_unref(&sync->sync_to[i]);
 
-	amdgpu_fence_unref(&sync->last_vm_update);
+	fence_put(sync->last_vm_update);
 }
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_test.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_test.c
index 962dd5552137..f80b1a43be8a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_test.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_test.c
@@ -77,7 +77,7 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev)
 		void *gtt_map, *vram_map;
 		void **gtt_start, **gtt_end;
 		void **vram_start, **vram_end;
-		struct amdgpu_fence *fence = NULL;
+		struct fence *fence = NULL;
 
 		r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
 				     AMDGPU_GEM_DOMAIN_GTT, 0, NULL, gtt_obj + i);
@@ -116,13 +116,13 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev)
 			goto out_lclean_unpin;
 		}
 
-		r = fence_wait(&fence->base, false);
+		r = fence_wait(fence, false);
 		if (r) {
 			DRM_ERROR("Failed to wait for GTT->VRAM fence %d\n", i);
 			goto out_lclean_unpin;
 		}
 
-		amdgpu_fence_unref(&fence);
+		fence_put(fence);
 
 		r = amdgpu_bo_kmap(vram_obj, &vram_map);
 		if (r) {
@@ -161,13 +161,13 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev)
 			goto out_lclean_unpin;
 		}
 
-		r = fence_wait(&fence->base, false);
+		r = fence_wait(fence, false);
 		if (r) {
 			DRM_ERROR("Failed to wait for VRAM->GTT fence %d\n", i);
 			goto out_lclean_unpin;
 		}
 
-		amdgpu_fence_unref(&fence);
+		fence_put(fence);
 
 		r = amdgpu_bo_kmap(gtt_obj[i], &gtt_map);
 		if (r) {
@@ -214,7 +214,7 @@ out_lclean:
 			amdgpu_bo_unref(&gtt_obj[i]);
 		}
 		if (fence)
-			amdgpu_fence_unref(&fence);
+			fence_put(fence);
 		break;
 	}
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index dd3415d2e45d..399143541d8a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -228,7 +228,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 	struct amdgpu_device *adev;
 	struct amdgpu_ring *ring;
 	uint64_t old_start, new_start;
-	struct amdgpu_fence *fence;
+	struct fence *fence;
 	int r;
 
 	adev = amdgpu_get_adev(bo->bdev);
@@ -269,9 +269,9 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 			       new_mem->num_pages * PAGE_SIZE, /* bytes */
 			       bo->resv, &fence);
 	/* FIXME: handle copy error */
-	r = ttm_bo_move_accel_cleanup(bo, &fence->base,
+	r = ttm_bo_move_accel_cleanup(bo, fence,
 				      evict, no_wait_gpu, new_mem);
-	amdgpu_fence_unref(&fence);
+	fence_put(fence);
 	return r;
 }
 
@@ -987,46 +987,48 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring,
 		       uint64_t dst_offset,
 		       uint32_t byte_count,
 		       struct reservation_object *resv,
-		       struct amdgpu_fence **fence)
+		       struct fence **fence)
 {
 	struct amdgpu_device *adev = ring->adev;
-	struct amdgpu_sync sync;
 	uint32_t max_bytes;
 	unsigned num_loops, num_dw;
+	struct amdgpu_ib *ib;
 	unsigned i;
 	int r;
 
-	/* sync other rings */
-	amdgpu_sync_create(&sync);
-	if (resv) {
-		r = amdgpu_sync_resv(adev, &sync, resv, false);
-		if (r) {
-			DRM_ERROR("sync failed (%d).\n", r);
-			amdgpu_sync_free(adev, &sync, NULL);
-			return r;
-		}
-	}
-
 	max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
 	num_loops = DIV_ROUND_UP(byte_count, max_bytes);
 	num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
 
-	/* for fence and sync */
-	num_dw += 64 + AMDGPU_NUM_SYNCS * 8;
+	/* for IB padding */
+	while (num_dw & 0x7)
+		num_dw++;
+
+	ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
+	if (!ib)
+		return -ENOMEM;
 
-	r = amdgpu_ring_lock(ring, num_dw);
+	r = amdgpu_ib_get(ring, NULL, num_dw * 4, ib);
 	if (r) {
-		DRM_ERROR("ring lock failed (%d).\n", r);
-		amdgpu_sync_free(adev, &sync, NULL);
+		kfree(ib);
 		return r;
 	}
 
-	amdgpu_sync_rings(&sync, ring);
+	ib->length_dw = 0;
+
+	if (resv) {
+		r = amdgpu_sync_resv(adev, &ib->sync, resv,
+				     AMDGPU_FENCE_OWNER_UNDEFINED);
+		if (r) {
+			DRM_ERROR("sync failed (%d).\n", r);
+			goto error_free;
+		}
+	}
 
 	for (i = 0; i < num_loops; i++) {
 		uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
 
-		amdgpu_emit_copy_buffer(adev, ring, src_offset, dst_offset,
+		amdgpu_emit_copy_buffer(adev, ib, src_offset, dst_offset,
 					cur_size_in_bytes);
 
 		src_offset += cur_size_in_bytes;
@@ -1034,17 +1036,24 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring,
 		byte_count -= cur_size_in_bytes;
 	}
 
-	r = amdgpu_fence_emit(ring, AMDGPU_FENCE_OWNER_MOVE, fence);
-	if (r) {
-		amdgpu_ring_unlock_undo(ring);
-		amdgpu_sync_free(adev, &sync, NULL);
-		return r;
-	}
-
-	amdgpu_ring_unlock_commit(ring);
-	amdgpu_sync_free(adev, &sync, *fence);
+	amdgpu_vm_pad_ib(adev, ib);
+	WARN_ON(ib->length_dw > num_dw);
+	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
+						 &amdgpu_vm_free_job,
+						 AMDGPU_FENCE_OWNER_MOVE,
+						 fence);
+	if (r)
+		goto error_free;
 
+	if (!amdgpu_enable_scheduler) {
+		amdgpu_ib_free(adev, ib);
+		kfree(ib);
+	}
 	return 0;
+error_free:
+	amdgpu_ib_free(adev, ib);
+	kfree(ib);
+	return r;
 }
 
 #if defined(CONFIG_DEBUG_FS)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
index 68369cf1e318..b87355ccfb1d 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
@@ -807,7 +807,7 @@ int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
 }
 
 static int amdgpu_uvd_free_job(
-	struct amdgpu_cs_parser *sched_job)
+	struct amdgpu_job *sched_job)
 {
 	amdgpu_ib_free(sched_job->adev, sched_job->ibs);
 	kfree(sched_job->ibs);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
index 33ee6ae28f37..1a984c934b1f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
@@ -340,7 +340,7 @@ void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
 }
 
 static int amdgpu_vce_free_job(
-	struct amdgpu_cs_parser *sched_job)
+	struct amdgpu_job *sched_job)
 {
 	amdgpu_ib_free(sched_job->adev, sched_job->ibs);
 	kfree(sched_job->ibs);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index a78a206e176e..83b7ce6f5f72 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -200,19 +200,29 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
  */
 void amdgpu_vm_flush(struct amdgpu_ring *ring,
 		     struct amdgpu_vm *vm,
-		     struct amdgpu_fence *updates)
+		     struct fence *updates)
 {
 	uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
 	struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
-	struct amdgpu_fence *flushed_updates = vm_id->flushed_updates;
+	struct fence *flushed_updates = vm_id->flushed_updates;
+	bool is_earlier = false;
+
+	if (flushed_updates && updates) {
+		BUG_ON(flushed_updates->context != updates->context);
+		is_earlier = (updates->seqno - flushed_updates->seqno <=
+			      INT_MAX) ? true : false;
+	}
 
 	if (pd_addr != vm_id->pd_gpu_addr || !flushed_updates ||
-	    (updates && amdgpu_fence_is_earlier(flushed_updates, updates))) {
+	    is_earlier) {
 
 		trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id);
-		vm_id->flushed_updates = amdgpu_fence_ref(
-			amdgpu_fence_later(flushed_updates, updates));
-		amdgpu_fence_unref(&flushed_updates);
+		if (is_earlier) {
+			vm_id->flushed_updates = fence_get(updates);
+			fence_put(flushed_updates);
+		}
+		if (!flushed_updates)
+			vm_id->flushed_updates = fence_get(updates);
 		vm_id->pd_gpu_addr = pd_addr;
 		amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr);
 	}
@@ -306,8 +316,7 @@ static void amdgpu_vm_update_pages(struct amdgpu_device *adev,
 	}
 }
 
-static int amdgpu_vm_free_job(
-	struct amdgpu_cs_parser *sched_job)
+int amdgpu_vm_free_job(struct amdgpu_job *sched_job)
 {
 	int i;
 	for (i = 0; i < sched_job->num_ibs; i++)
@@ -1347,7 +1356,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	fence_put(vm->page_directory_fence);
 
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-		amdgpu_fence_unref(&vm->ids[i].flushed_updates);
+		fence_put(vm->ids[i].flushed_updates);
 		amdgpu_fence_unref(&vm->ids[i].last_id_use);
 	}
 
diff --git a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c
index 2b4242b39b0a..3920c1e346f8 100644
--- a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c
+++ b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c
@@ -630,6 +630,7 @@ static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring)
 	gpu_addr = adev->wb.gpu_addr + (index * 4);
 	tmp = 0xCAFEDEAD;
 	adev->wb.wb[index] = cpu_to_le32(tmp);
+	memset(&ib, 0, sizeof(ib));
 	r = amdgpu_ib_get(ring, NULL, 256, &ib);
 	if (r) {
 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
@@ -1338,18 +1339,18 @@ static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev)
  * Used by the amdgpu ttm implementation to move pages if
  * registered as the asic copy callback.
  */
-static void cik_sdma_emit_copy_buffer(struct amdgpu_ring *ring,
+static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib,
 				      uint64_t src_offset,
 				      uint64_t dst_offset,
 				      uint32_t byte_count)
 {
-	amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0));
-	amdgpu_ring_write(ring, byte_count);
-	amdgpu_ring_write(ring, 0); /* src/dst endian swap */
-	amdgpu_ring_write(ring, lower_32_bits(src_offset));
-	amdgpu_ring_write(ring, upper_32_bits(src_offset));
-	amdgpu_ring_write(ring, lower_32_bits(dst_offset));
-	amdgpu_ring_write(ring, upper_32_bits(dst_offset));
+	ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0);
+	ib->ptr[ib->length_dw++] = byte_count;
+	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
+	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
+	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
+	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
+	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
 }
 
 /**
diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c
index 9b0cab413677..fab7b236f37f 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c
@@ -2660,6 +2660,7 @@ static int gfx_v7_0_ring_test_ib(struct amdgpu_ring *ring)
 		return r;
 	}
 	WREG32(scratch, 0xCAFEDEAD);
+	memset(&ib, 0, sizeof(ib));
 	r = amdgpu_ib_get(ring, NULL, 256, &ib);
 	if (r) {
 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
index 4b68e6306f40..818edb37fa9c 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
@@ -622,6 +622,7 @@ static int gfx_v8_0_ring_test_ib(struct amdgpu_ring *ring)
 		return r;
 	}
 	WREG32(scratch, 0xCAFEDEAD);
+	memset(&ib, 0, sizeof(ib));
 	r = amdgpu_ib_get(ring, NULL, 256, &ib);
 	if (r) {
 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c b/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c
index 9de8104eddeb..715e02d3bfba 100644
--- a/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c
+++ b/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c
@@ -689,6 +689,7 @@ static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring)
 	gpu_addr = adev->wb.gpu_addr + (index * 4);
 	tmp = 0xCAFEDEAD;
 	adev->wb.wb[index] = cpu_to_le32(tmp);
+	memset(&ib, 0, sizeof(ib));
 	r = amdgpu_ib_get(ring, NULL, 256, &ib);
 	if (r) {
 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
@@ -1349,19 +1350,19 @@ static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev)
  * Used by the amdgpu ttm implementation to move pages if
  * registered as the asic copy callback.
  */
-static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ring *ring,
+static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib,
 				       uint64_t src_offset,
 				       uint64_t dst_offset,
 				       uint32_t byte_count)
 {
-	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
-			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR));
-	amdgpu_ring_write(ring, byte_count);
-	amdgpu_ring_write(ring, 0); /* src/dst endian swap */
-	amdgpu_ring_write(ring, lower_32_bits(src_offset));
-	amdgpu_ring_write(ring, upper_32_bits(src_offset));
-	amdgpu_ring_write(ring, lower_32_bits(dst_offset));
-	amdgpu_ring_write(ring, upper_32_bits(dst_offset));
+	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
+		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
+	ib->ptr[ib->length_dw++] = byte_count;
+	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
+	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
+	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
+	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
+	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
 }
 
 /**
diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c
index 029f3455f9f9..67128c8e78b8 100644
--- a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c
@@ -810,6 +810,7 @@ static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring)
 	gpu_addr = adev->wb.gpu_addr + (index * 4);
 	tmp = 0xCAFEDEAD;
 	adev->wb.wb[index] = cpu_to_le32(tmp);
+	memset(&ib, 0, sizeof(ib));
 	r = amdgpu_ib_get(ring, NULL, 256, &ib);
 	if (r) {
 		DRM_ERROR("amdgpu: failed to get ib (%d).\n", r);
@@ -1473,19 +1474,19 @@ static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev)
  * Used by the amdgpu ttm implementation to move pages if
  * registered as the asic copy callback.
  */
-static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ring *ring,
+static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ib *ib,
 				       uint64_t src_offset,
 				       uint64_t dst_offset,
 				       uint32_t byte_count)
 {
-	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
-			  SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR));
-	amdgpu_ring_write(ring, byte_count);
-	amdgpu_ring_write(ring, 0); /* src/dst endian swap */
-	amdgpu_ring_write(ring, lower_32_bits(src_offset));
-	amdgpu_ring_write(ring, upper_32_bits(src_offset));
-	amdgpu_ring_write(ring, lower_32_bits(dst_offset));
-	amdgpu_ring_write(ring, upper_32_bits(dst_offset));
+	ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) |
+		SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR);
+	ib->ptr[ib->length_dw++] = byte_count;
+	ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */
+	ib->ptr[ib->length_dw++] = lower_32_bits(src_offset);
+	ib->ptr[ib->length_dw++] = upper_32_bits(src_offset);
+	ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset);
+	ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset);
 }
 
 /**
diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
index 265d3e2f63cc..d99fe90991dc 100644
--- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
+++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
@@ -27,30 +27,32 @@
 #include <drm/drmP.h>
 #include "gpu_scheduler.h"
 
+static void amd_sched_wakeup(struct amd_gpu_scheduler *sched);
+
 /* Initialize a given run queue struct */
 static void amd_sched_rq_init(struct amd_sched_rq *rq)
 {
+	spin_lock_init(&rq->lock);
 	INIT_LIST_HEAD(&rq->entities);
-	mutex_init(&rq->lock);
 	rq->current_entity = NULL;
 }
 
 static void amd_sched_rq_add_entity(struct amd_sched_rq *rq,
 				    struct amd_sched_entity *entity)
 {
-	mutex_lock(&rq->lock);
+	spin_lock(&rq->lock);
 	list_add_tail(&entity->list, &rq->entities);
-	mutex_unlock(&rq->lock);
+	spin_unlock(&rq->lock);
 }
 
 static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
 				       struct amd_sched_entity *entity)
 {
-	mutex_lock(&rq->lock);
+	spin_lock(&rq->lock);
 	list_del_init(&entity->list);
 	if (rq->current_entity == entity)
 		rq->current_entity = NULL;
-	mutex_unlock(&rq->lock);
+	spin_unlock(&rq->lock);
 }
 
 /**
@@ -61,12 +63,16 @@ static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq,
 static struct amd_sched_entity *
 amd_sched_rq_select_entity(struct amd_sched_rq *rq)
 {
-	struct amd_sched_entity *entity = rq->current_entity;
+	struct amd_sched_entity *entity;
+
+	spin_lock(&rq->lock);
 
+	entity = rq->current_entity;
 	if (entity) {
 		list_for_each_entry_continue(entity, &rq->entities, list) {
 			if (!kfifo_is_empty(&entity->job_queue)) {
 				rq->current_entity = entity;
+				spin_unlock(&rq->lock);
 				return rq->current_entity;
 			}
 		}
@@ -76,6 +82,7 @@ amd_sched_rq_select_entity(struct amd_sched_rq *rq)
 
 		if (!kfifo_is_empty(&entity->job_queue)) {
 			rq->current_entity = entity;
+			spin_unlock(&rq->lock);
 			return rq->current_entity;
 		}
 
@@ -83,76 +90,9 @@ amd_sched_rq_select_entity(struct amd_sched_rq *rq)
 			break;
 	}
 
-	return NULL;
-}
+	spin_unlock(&rq->lock);
 
-/**
- * Note: This function should only been called inside scheduler main
- * function for thread safety, there is no other protection here.
- * return ture if scheduler has something ready to run.
- *
- * For active_hw_rq, there is only one producer(scheduler thread) and
- * one consumer(ISR). It should be safe to use this function in scheduler
- * main thread to decide whether to continue emit more IBs.
-*/
-static bool is_scheduler_ready(struct amd_gpu_scheduler *sched)
-{
-	unsigned long flags;
-	bool full;
-
-	spin_lock_irqsave(&sched->queue_lock, flags);
-	full = atomic64_read(&sched->hw_rq_count) <
-		sched->hw_submission_limit ? true : false;
-	spin_unlock_irqrestore(&sched->queue_lock, flags);
-
-	return full;
-}
-
-/**
- * Select next entity from the kernel run queue, if not available,
- * return null.
-*/
-static struct amd_sched_entity *
-kernel_rq_select_context(struct amd_gpu_scheduler *sched)
-{
-	struct amd_sched_entity *sched_entity;
-	struct amd_sched_rq *rq = &sched->kernel_rq;
-
-	mutex_lock(&rq->lock);
-	sched_entity = amd_sched_rq_select_entity(rq);
-	mutex_unlock(&rq->lock);
-	return sched_entity;
-}
-
-/**
- * Select next entity containing real IB submissions
-*/
-static struct amd_sched_entity *
-select_context(struct amd_gpu_scheduler *sched)
-{
-	struct amd_sched_entity *wake_entity = NULL;
-	struct amd_sched_entity *tmp;
-	struct amd_sched_rq *rq;
-
-	if (!is_scheduler_ready(sched))
-		return NULL;
-
-	/* Kernel run queue has higher priority than normal run queue*/
-	tmp = kernel_rq_select_context(sched);
-	if (tmp != NULL)
-		goto exit;
-
-	rq = &sched->sched_rq;
-	mutex_lock(&rq->lock);
-	tmp = amd_sched_rq_select_entity(rq);
-	mutex_unlock(&rq->lock);
-exit:
-	if (sched->current_entity && (sched->current_entity != tmp))
-		wake_entity = sched->current_entity;
-	sched->current_entity = tmp;
-	if (wake_entity && wake_entity->need_wakeup)
-		wake_up(&wake_entity->wait_queue);
-	return tmp;
+	return NULL;
 }
 
 /**
@@ -171,31 +111,20 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 			  struct amd_sched_rq *rq,
 			  uint32_t jobs)
 {
-	uint64_t seq_ring = 0;
-	char name[20];
-
 	if (!(sched && entity && rq))
 		return -EINVAL;
 
 	memset(entity, 0, sizeof(struct amd_sched_entity));
-	seq_ring = ((uint64_t)sched->ring_id) << 60;
-	spin_lock_init(&entity->lock);
 	entity->belongto_rq = rq;
 	entity->scheduler = sched;
-	init_waitqueue_head(&entity->wait_queue);
-	init_waitqueue_head(&entity->wait_emit);
 	entity->fence_context = fence_context_alloc(1);
-	snprintf(name, sizeof(name), "c_entity[%llu]", entity->fence_context);
-	memcpy(entity->name, name, 20);
-	entity->need_wakeup = false;
 	if(kfifo_alloc(&entity->job_queue,
 		       jobs * sizeof(void *),
 		       GFP_KERNEL))
 		return -EINVAL;
 
 	spin_lock_init(&entity->queue_lock);
-	atomic64_set(&entity->last_queued_v_seq, seq_ring);
-	atomic64_set(&entity->last_signaled_v_seq, seq_ring);
+	atomic_set(&entity->fence_seq, 0);
 
 	/* Add the entity to the run queue */
 	amd_sched_rq_add_entity(rq, entity);
@@ -210,23 +139,24 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
  *
  * return true if entity is initialized, false otherwise
 */
-static bool is_context_entity_initialized(struct amd_gpu_scheduler *sched,
-					  struct amd_sched_entity *entity)
+static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched,
+					    struct amd_sched_entity *entity)
 {
 	return entity->scheduler == sched &&
 		entity->belongto_rq != NULL;
 }
 
-static bool is_context_entity_idle(struct amd_gpu_scheduler *sched,
-				   struct amd_sched_entity *entity)
+/**
+ * Check if entity is idle
+ *
+ * @entity	The pointer to a valid scheduler entity
+ *
+ * Return true if entity don't has any unscheduled jobs.
+ */
+static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity)
 {
-	/**
-	 * Idle means no pending IBs, and the entity is not
-	 * currently being used.
-	*/
-	barrier();
-	if ((sched->current_entity != entity) &&
-	    kfifo_is_empty(&entity->job_queue))
+	rmb();
+	if (kfifo_is_empty(&entity->job_queue))
 		return true;
 
 	return false;
@@ -238,84 +168,114 @@ static bool is_context_entity_idle(struct amd_gpu_scheduler *sched,
  * @sched       Pointer to scheduler instance
  * @entity	The pointer to a valid scheduler entity
  *
- * return 0 if succeed. negative error code on failure
+ * Cleanup and free the allocated resources.
  */
-int amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
-			    struct amd_sched_entity *entity)
+void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
+			   struct amd_sched_entity *entity)
 {
-	int r = 0;
 	struct amd_sched_rq *rq = entity->belongto_rq;
 
-	if (!is_context_entity_initialized(sched, entity))
-		return 0;
-	entity->need_wakeup = true;
+	if (!amd_sched_entity_is_initialized(sched, entity))
+		return;
+
 	/**
 	 * The client will not queue more IBs during this fini, consume existing
 	 * queued IBs
 	*/
-	r = wait_event_timeout(
-		entity->wait_queue,
-		is_context_entity_idle(sched, entity),
-		msecs_to_jiffies(AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS)
-		) ?  0 : -1;
-
-	if (r) {
-		if (entity->is_pending)
-			DRM_INFO("Entity %p is in waiting state during fini,\
-				all pending ibs will be canceled.\n",
-				 entity);
-	}
+	wait_event(sched->job_scheduled, amd_sched_entity_is_idle(entity));
 
 	amd_sched_rq_remove_entity(rq, entity);
 	kfifo_free(&entity->job_queue);
-	return r;
 }
 
 /**
- * Submit a normal job to the job queue
+ * Helper to submit a job to the job queue
  *
- * @sched	The pointer to the scheduler
- * @c_entity    The pointer to amd_sched_entity
  * @job		The pointer to job required to submit
- * return 0 if succeed. -1 if failed.
- *        -2 indicate queue is full for this client, client should wait untill
- *	     scheduler consum some queued command.
- *	  -1 other fail.
-*/
-int amd_sched_push_job(struct amd_gpu_scheduler *sched,
-		       struct amd_sched_entity *c_entity,
-		       void *data,
-		       struct amd_sched_fence **fence)
+ *
+ * Returns true if we could submit the job.
+ */
+static bool amd_sched_entity_in(struct amd_sched_job *job)
+{
+	struct amd_sched_entity *entity = job->s_entity;
+	bool added, first = false;
+
+	spin_lock(&entity->queue_lock);
+	added = kfifo_in(&entity->job_queue, &job, sizeof(job)) == sizeof(job);
+
+	if (added && kfifo_len(&entity->job_queue) == sizeof(job))
+		first = true;
+
+	spin_unlock(&entity->queue_lock);
+
+	/* first job wakes up scheduler */
+	if (first)
+		amd_sched_wakeup(job->sched);
+
+	return added;
+}
+
+/**
+ * Submit a job to the job queue
+ *
+ * @job		The pointer to job required to submit
+ *
+ * Returns 0 for success, negative error code otherwise.
+ */
+int amd_sched_entity_push_job(struct amd_sched_job *sched_job)
 {
-	struct amd_sched_job *job;
+	struct amd_sched_entity *entity = sched_job->s_entity;
+	struct amd_sched_fence *fence = amd_sched_fence_create(
+		entity, sched_job->owner);
+	int r;
 
 	if (!fence)
-		return -EINVAL;
-	job = kzalloc(sizeof(struct amd_sched_job), GFP_KERNEL);
-	if (!job)
 		return -ENOMEM;
-	job->sched = sched;
-	job->s_entity = c_entity;
-	job->data = data;
-	*fence = amd_sched_fence_create(c_entity);
-	if ((*fence) == NULL) {
-		kfree(job);
-		return -EINVAL;
-	}
-	fence_get(&(*fence)->base);
-	job->s_fence = *fence;
-	while (kfifo_in_spinlocked(&c_entity->job_queue, &job, sizeof(void *),
-				   &c_entity->queue_lock) != sizeof(void *)) {
-		/**
-		 * Current context used up all its IB slots
-		 * wait here, or need to check whether GPU is hung
-		*/
-		schedule();
-	}
-	/* first job wake up scheduler */
-	if ((kfifo_len(&c_entity->job_queue) / sizeof(void *)) == 1)
-		wake_up_interruptible(&sched->wait_queue);
-	return 0;
+
+	fence_get(&fence->base);
+	sched_job->s_fence = fence;
+
+	r = wait_event_interruptible(entity->scheduler->job_scheduled,
+				     amd_sched_entity_in(sched_job));
+
+	return r;
+}
+
+/**
+ * Return ture if we can push more jobs to the hw.
+ */
+static bool amd_sched_ready(struct amd_gpu_scheduler *sched)
+{
+	return atomic_read(&sched->hw_rq_count) <
+		sched->hw_submission_limit;
+}
+
+/**
+ * Wake up the scheduler when it is ready
+ */
+static void amd_sched_wakeup(struct amd_gpu_scheduler *sched)
+{
+	if (amd_sched_ready(sched))
+		wake_up_interruptible(&sched->wake_up_worker);
+}
+
+/**
+ * Select next entity containing real IB submissions
+*/
+static struct amd_sched_entity *
+amd_sched_select_context(struct amd_gpu_scheduler *sched)
+{
+	struct amd_sched_entity *tmp;
+
+	if (!amd_sched_ready(sched))
+		return NULL;
+
+	/* Kernel run queue has higher priority than normal run queue*/
+	tmp = amd_sched_rq_select_entity(&sched->kernel_rq);
+	if (tmp == NULL)
+		tmp = amd_sched_rq_select_entity(&sched->sched_rq);
+
+	return tmp;
 }
 
 static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
@@ -323,52 +283,41 @@ static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 	struct amd_sched_job *sched_job =
 		container_of(cb, struct amd_sched_job, cb);
 	struct amd_gpu_scheduler *sched;
-	unsigned long flags;
 
 	sched = sched_job->sched;
-	atomic64_set(&sched_job->s_entity->last_signaled_v_seq,
-		     sched_job->s_fence->v_seq);
 	amd_sched_fence_signal(sched_job->s_fence);
-	spin_lock_irqsave(&sched->queue_lock, flags);
-	list_del(&sched_job->list);
-	atomic64_dec(&sched->hw_rq_count);
-	spin_unlock_irqrestore(&sched->queue_lock, flags);
-
-	sched->ops->process_job(sched, sched_job);
+	atomic_dec(&sched->hw_rq_count);
 	fence_put(&sched_job->s_fence->base);
-	kfree(sched_job);
-	wake_up_interruptible(&sched->wait_queue);
+	sched->ops->process_job(sched_job);
+	wake_up_interruptible(&sched->wake_up_worker);
 }
 
 static int amd_sched_main(void *param)
 {
-	int r;
-	struct amd_sched_job *job;
 	struct sched_param sparam = {.sched_priority = 1};
-	struct amd_sched_entity *c_entity = NULL;
 	struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param;
+	int r;
 
 	sched_setscheduler(current, SCHED_FIFO, &sparam);
 
 	while (!kthread_should_stop()) {
+		struct amd_sched_entity *c_entity = NULL;
+		struct amd_sched_job *job;
 		struct fence *fence;
 
-		wait_event_interruptible(sched->wait_queue,
-					 is_scheduler_ready(sched) &&
-					 (c_entity = select_context(sched)));
+		wait_event_interruptible(sched->wake_up_worker,
+			kthread_should_stop() ||
+			(c_entity = amd_sched_select_context(sched)));
+
+		if (!c_entity)
+			continue;
+
 		r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *));
 		if (r != sizeof(void *))
 			continue;
-		r = sched->ops->prepare_job(sched, c_entity, job);
-		if (!r) {
-			unsigned long flags;
-			spin_lock_irqsave(&sched->queue_lock, flags);
-			list_add_tail(&job->list, &sched->active_hw_rq);
-			atomic64_inc(&sched->hw_rq_count);
-			spin_unlock_irqrestore(&sched->queue_lock, flags);
-		}
-		mutex_lock(&sched->sched_lock);
-		fence = sched->ops->run_job(sched, c_entity, job);
+		atomic_inc(&sched->hw_rq_count);
+
+		fence = sched->ops->run_job(job);
 		if (fence) {
 			r = fence_add_callback(fence, &job->cb,
 					       amd_sched_process_job);
@@ -378,7 +327,8 @@ static int amd_sched_main(void *param)
 				DRM_ERROR("fence add callback failed (%d)\n", r);
 			fence_put(fence);
 		}
-		mutex_unlock(&sched->sched_lock);
+
+		wake_up(&sched->job_scheduled);
 	}
 	return 0;
 }
@@ -386,53 +336,42 @@ static int amd_sched_main(void *param)
 /**
  * Create a gpu scheduler
  *
- * @device	The device context for this scheduler
- * @ops		The backend operations for this scheduler.
- * @id	        The scheduler is per ring, here is ring id.
- * @granularity	The minumum ms unit the scheduler will scheduled.
- * @preemption  Indicate whether this ring support preemption, 0 is no.
+ * @ops			The backend operations for this scheduler.
+ * @ring		The the ring id for the scheduler.
+ * @hw_submissions	Number of hw submissions to do.
  *
- * return the pointer to scheduler for success, otherwise return NULL
+ * Return the pointer to scheduler for success, otherwise return NULL
 */
-struct amd_gpu_scheduler *amd_sched_create(void *device,
-					   struct amd_sched_backend_ops *ops,
-					   unsigned ring,
-					   unsigned granularity,
-					   unsigned preemption,
-					   unsigned hw_submission)
+struct amd_gpu_scheduler *amd_sched_create(struct amd_sched_backend_ops *ops,
+					   unsigned ring, unsigned hw_submission,
+					   void *priv)
 {
 	struct amd_gpu_scheduler *sched;
-	char name[20];
 
 	sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL);
 	if (!sched)
 		return NULL;
 
-	sched->device = device;
 	sched->ops = ops;
-	sched->granularity = granularity;
 	sched->ring_id = ring;
-	sched->preemption = preemption;
 	sched->hw_submission_limit = hw_submission;
-	snprintf(name, sizeof(name), "gpu_sched[%d]", ring);
-	mutex_init(&sched->sched_lock);
-	spin_lock_init(&sched->queue_lock);
+	sched->priv = priv;
+	snprintf(sched->name, sizeof(sched->name), "amdgpu[%d]", ring);
 	amd_sched_rq_init(&sched->sched_rq);
 	amd_sched_rq_init(&sched->kernel_rq);
 
-	init_waitqueue_head(&sched->wait_queue);
-	INIT_LIST_HEAD(&sched->active_hw_rq);
-	atomic64_set(&sched->hw_rq_count, 0);
+	init_waitqueue_head(&sched->wake_up_worker);
+	init_waitqueue_head(&sched->job_scheduled);
+	atomic_set(&sched->hw_rq_count, 0);
 	/* Each scheduler will run on a seperate kernel thread */
-	sched->thread = kthread_create(amd_sched_main, sched, name);
-	if (sched->thread) {
-		wake_up_process(sched->thread);
-		return sched;
+	sched->thread = kthread_run(amd_sched_main, sched, sched->name);
+	if (IS_ERR(sched->thread)) {
+		DRM_ERROR("Failed to create scheduler for id %d.\n", ring);
+		kfree(sched);
+		return NULL;
 	}
 
-	DRM_ERROR("Failed to create scheduler for id %d.\n", ring);
-	kfree(sched);
-	return NULL;
+	return sched;
 }
 
 /**
@@ -448,15 +387,3 @@ int amd_sched_destroy(struct amd_gpu_scheduler *sched)
 	kfree(sched);
 	return  0;
 }
-
-/**
- * Get next queued sequence number
- *
- * @entity The context entity
- *
- * return the next queued sequence number
-*/
-uint64_t amd_sched_next_queued_seq(struct amd_sched_entity *c_entity)
-{
-	return atomic64_read(&c_entity->last_queued_v_seq) + 1;
-}
diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
index ceb5918bfbeb..e797796dcad7 100644
--- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
+++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
@@ -27,8 +27,6 @@
 #include <linux/kfifo.h>
 #include <linux/fence.h>
 
-#define AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS		3000
-
 struct amd_gpu_scheduler;
 struct amd_sched_rq;
 
@@ -41,20 +39,12 @@ struct amd_sched_rq;
 struct amd_sched_entity {
 	struct list_head		list;
 	struct amd_sched_rq		*belongto_rq;
-	spinlock_t			lock;
-	/* the virtual_seq is unique per context per ring */
-	atomic64_t			last_queued_v_seq;
-	atomic64_t			last_signaled_v_seq;
+	atomic_t			fence_seq;
 	/* the job_queue maintains the jobs submitted by clients */
 	struct kfifo                    job_queue;
 	spinlock_t			queue_lock;
 	struct amd_gpu_scheduler	*scheduler;
-	wait_queue_head_t		wait_queue;
-	wait_queue_head_t		wait_emit;
-	bool                            is_pending;
 	uint64_t                        fence_context;
-	char                            name[20];
-	bool                            need_wakeup;
 };
 
 /**
@@ -63,26 +53,24 @@ struct amd_sched_entity {
  * the next entity to emit commands from.
 */
 struct amd_sched_rq {
-	struct mutex		lock;
+	spinlock_t		lock;
 	struct list_head	entities;
 	struct amd_sched_entity	*current_entity;
 };
 
 struct amd_sched_fence {
 	struct fence                    base;
-	struct fence_cb                 cb;
-	struct amd_sched_entity	        *entity;
-	uint64_t			v_seq;
+	struct amd_gpu_scheduler	*scheduler;
 	spinlock_t			lock;
+	void                            *owner;
 };
 
 struct amd_sched_job {
-	struct list_head		list;
 	struct fence_cb                 cb;
 	struct amd_gpu_scheduler        *sched;
 	struct amd_sched_entity         *s_entity;
-	void                            *data;
 	struct amd_sched_fence          *s_fence;
+	void		                *owner;
 };
 
 extern const struct fence_ops amd_sched_fence_ops;
@@ -101,61 +89,42 @@ static inline struct amd_sched_fence *to_amd_sched_fence(struct fence *f)
  * these functions should be implemented in driver side
 */
 struct amd_sched_backend_ops {
-	int (*prepare_job)(struct amd_gpu_scheduler *sched,
-			   struct amd_sched_entity *c_entity,
-			   struct amd_sched_job *job);
-	struct fence *(*run_job)(struct amd_gpu_scheduler *sched,
-				 struct amd_sched_entity *c_entity,
-				 struct amd_sched_job *job);
-	void (*process_job)(struct amd_gpu_scheduler *sched,
-				    struct amd_sched_job *job);
+	struct fence *(*run_job)(struct amd_sched_job *job);
+	void (*process_job)(struct amd_sched_job *job);
 };
 
 /**
  * One scheduler is implemented for each hardware ring
 */
 struct amd_gpu_scheduler {
-	void			        *device;
 	struct task_struct		*thread;
 	struct amd_sched_rq		sched_rq;
 	struct amd_sched_rq		kernel_rq;
-	struct list_head		active_hw_rq;
-	atomic64_t			hw_rq_count;
+	atomic_t			hw_rq_count;
 	struct amd_sched_backend_ops	*ops;
 	uint32_t			ring_id;
-	uint32_t			granularity; /* in ms unit */
-	uint32_t			preemption;
-	wait_queue_head_t		wait_queue;
-	struct amd_sched_entity	*current_entity;
-	struct mutex			sched_lock;
-	spinlock_t			queue_lock;
+	wait_queue_head_t		wake_up_worker;
+	wait_queue_head_t		job_scheduled;
 	uint32_t                        hw_submission_limit;
+	char                            name[20];
+	void                            *priv;
 };
 
-struct amd_gpu_scheduler *amd_sched_create(void *device,
-				struct amd_sched_backend_ops *ops,
-				uint32_t ring,
-				uint32_t granularity,
-				uint32_t preemption,
-				uint32_t hw_submission);
+struct amd_gpu_scheduler *
+amd_sched_create(struct amd_sched_backend_ops *ops,
+		 uint32_t ring, uint32_t hw_submission, void *priv);
 int amd_sched_destroy(struct amd_gpu_scheduler *sched);
 
-int amd_sched_push_job(struct amd_gpu_scheduler *sched,
-		       struct amd_sched_entity *c_entity,
-		       void *data,
-		       struct amd_sched_fence **fence);
-
 int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 			  struct amd_sched_entity *entity,
 			  struct amd_sched_rq *rq,
 			  uint32_t jobs);
-int amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
-			  struct amd_sched_entity *entity);
-
-uint64_t amd_sched_next_queued_seq(struct amd_sched_entity *c_entity);
+void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
+			   struct amd_sched_entity *entity);
+int amd_sched_entity_push_job(struct amd_sched_job *sched_job);
 
 struct amd_sched_fence *amd_sched_fence_create(
-	struct amd_sched_entity *s_entity);
+	struct amd_sched_entity *s_entity, void *owner);
 void amd_sched_fence_signal(struct amd_sched_fence *fence);
 
 
diff --git a/drivers/gpu/drm/amd/scheduler/sched_fence.c b/drivers/gpu/drm/amd/scheduler/sched_fence.c
index a4751598c0b4..e62c37920e11 100644
--- a/drivers/gpu/drm/amd/scheduler/sched_fence.c
+++ b/drivers/gpu/drm/amd/scheduler/sched_fence.c
@@ -27,19 +27,22 @@
 #include <drm/drmP.h>
 #include "gpu_scheduler.h"
 
-struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *s_entity)
+struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *s_entity, void *owner)
 {
 	struct amd_sched_fence *fence = NULL;
+	unsigned seq;
+
 	fence = kzalloc(sizeof(struct amd_sched_fence), GFP_KERNEL);
 	if (fence == NULL)
 		return NULL;
-	fence->v_seq = atomic64_inc_return(&s_entity->last_queued_v_seq);
-	fence->entity = s_entity;
+	fence->owner = owner;
+	fence->scheduler = s_entity->scheduler;
 	spin_lock_init(&fence->lock);
-	fence_init(&fence->base, &amd_sched_fence_ops,
-		&fence->lock,
-		s_entity->fence_context,
-		fence->v_seq);
+
+	seq = atomic_inc_return(&s_entity->fence_seq);
+	fence_init(&fence->base, &amd_sched_fence_ops, &fence->lock,
+		   s_entity->fence_context, seq);
+
 	return fence;
 }
 
@@ -60,7 +63,7 @@ static const char *amd_sched_fence_get_driver_name(struct fence *fence)
 static const char *amd_sched_fence_get_timeline_name(struct fence *f)
 {
 	struct amd_sched_fence *fence = to_amd_sched_fence(f);
-	return (const char *)fence->entity->name;
+	return (const char *)fence->scheduler->name;
 }
 
 static bool amd_sched_fence_enable_signaling(struct fence *f)
diff --git a/drivers/gpu/drm/radeon/radeon_connectors.c b/drivers/gpu/drm/radeon/radeon_connectors.c
index 94b21ae70ef7..5a2cafb4f1bc 100644
--- a/drivers/gpu/drm/radeon/radeon_connectors.c
+++ b/drivers/gpu/drm/radeon/radeon_connectors.c
@@ -95,6 +95,11 @@ void radeon_connector_hotplug(struct drm_connector *connector)
 			if (!radeon_hpd_sense(rdev, radeon_connector->hpd.hpd)) {
 				drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
 			} else if (radeon_dp_needs_link_train(radeon_connector)) {
+				/* Don't try to start link training before we
+				 * have the dpcd */
+				if (!radeon_dp_getdpcd(radeon_connector))
+					return;
+
 				/* set it to OFF so that drm_helper_connector_dpms()
 				 * won't return immediately since the current state
 				 * is ON at this point.