Merge tag 'drm-for-v4.8' of git://people.freedesktop.org/~airlied/linux

Merge drm updates from Dave Airlie:
 "This is the main drm pull request for 4.8.

  I'm down with a cold at the moment so hopefully this isn't in too bad
  a state, I finished pulling stuff last week mostly (nouveau fixes just
  went in today), so only this message should be influenced by illness.
  Apologies to anyone who's major feature I missed :-)

  Core:
        Lockless GEM BO freeing
        Non-blocking atomic work
        Documentation changes (rst/sphinx)
        Prep for new fencing changes
        Simple display helpers
        Master/auth changes
        Register/unregister rework
        Loads of trivial patches/fixes.

  New stuff:
        ARM Mali display driver (not the 3D chip)
        sii902x RGB->HDMI bridge

  Panel:
        Support for new panels
        Improved backlight support

  Bridge:
        Convert ADV7511 to bridge driver
        ADV7533 support
        TC358767 (DSI/DPI to eDP) encoder chip support

  i915:
        BXT support enabled by default
        GVT-g infrastructure
        GuC command submission and fixes
        BXT workarounds
        SKL/BKL workarounds
        Demidlayering device registration
        Thundering herd fixes
        Missing pci ids
        Atomic updates

  amdgpu/radeon:
        ATPX improvements for better dGPU power control on PX systems
        New power features for CZ/BR/ST
        Pipelined BO moves and evictions in TTM
        GPU scheduler improvements
        GPU reset improvements
        Overclocking on dGPUs with amdgpu
        Polaris powermanagement enabled

  nouveau:
        GK20A/GM20B volt and clock improvements.
        Initial support for GP100/GP104 GPUs, GP104 will not yet support
        acceleration due to NVIDIA having not released firmware for them as of yet.

  exynos:
        Exynos5433 SoC with IOMMU support.

  vc4:
        Shader validation for branching

  imx-drm:
        Atomic mode setting conversion
        Reworked DMFC FIFO allocation
        External bridge support

  analogix-dp:
        RK3399 eDP support
        Lots of fixes.

  rockchip:
        Lots of small fixes.

  msm:
        DT bindings cleanups
        Shrinker and madvise support
        ASoC HDMI codec support

  tegra:
        Host1x driver cleanups
        SOR reworking for DP support
        Runtime PM support

  omapdrm:
        PLL enhancements
        Header refactoring
        Gamma table support

  arcgpu:
        Simulator support

  virtio-gpu:
        Atomic modesetting fixes.

  rcar-du:
        Misc fixes.

  mediatek:
        MT8173 HDMI support

  sti:
        ASOC HDMI codec support
        Minor fixes

  fsl-dcu:
        Suspend/resume support
        Bridge support

  amdkfd:
        Minor fixes.

  etnaviv:
        Enable GPU clock gating

  hisilicon:
        Vblank and other fixes"

* tag 'drm-for-v4.8' of git://people.freedesktop.org/~airlied/linux: (1575 commits)
  drm/nouveau/gr/nv3x: fix instobj write offsets in gr setup
  drm/nouveau/acpi: fix lockup with PCIe runtime PM
  drm/nouveau/acpi: check for function 0x1B before using it
  drm/nouveau/acpi: return supported DSM functions
  drm/nouveau/acpi: ensure matching ACPI handle and supported functions
  drm/nouveau/fbcon: fix font width not divisible by 8
  drm/amd/powerplay: remove enable_clock_power_gatings_tasks from initialize and resume events
  drm/amd/powerplay: move clockgating to after ungating power in pp for uvd/vce
  drm/amdgpu: add query device id and revision id into system info entry at CGS
  drm/amdgpu: add new definition in bif header
  drm/amd/powerplay: rename smum header guards
  drm/amdgpu: enable UVD context buffer for older HW
  drm/amdgpu: fix default UVD context size
  drm/amdgpu: fix incorrect type of info_id
  drm/amdgpu: make amdgpu_cgs_call_acpi_method as static
  drm/amdgpu: comment out unused defaults_staturn_pro static const structure to fix the build
  drm/amdgpu: enable UVD VM only on polaris
  drm/amdgpu: increase timeout of IB test
  drm/amdgpu: add destroy session when generate VCE destroy msg.
  drm/amd: fix deadlock of job_list_lock V2
  ...

4 files changed, 204 insertions, 131 deletions

diff --git a/drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h b/drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h
index c89dc777768f..b961a1c6caf3 100644
--- a/drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h
+++ b/drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h
@@ -26,7 +26,7 @@ TRACE_EVENT(amd_sched_job,
 	    TP_fast_assign(
 			   __entry->entity = sched_job->s_entity;
 			   __entry->sched_job = sched_job;
-			   __entry->fence = &sched_job->s_fence->base;
+			   __entry->fence = &sched_job->s_fence->finished;
 			   __entry->name = sched_job->sched->name;
 			   __entry->job_count = kfifo_len(
 				   &sched_job->s_entity->job_queue) / sizeof(sched_job);
@@ -46,7 +46,7 @@ TRACE_EVENT(amd_sched_process_job,
 		    ),
 
 	    TP_fast_assign(
-		    __entry->fence = &fence->base;
+		    __entry->fence = &fence->finished;
 		    ),
 	    TP_printk("fence=%p signaled", __entry->fence)
 );
diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
index c16248cee779..ef312bb75fda 100644
--- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
+++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
@@ -32,6 +32,7 @@
 
 static bool amd_sched_entity_is_ready(struct amd_sched_entity *entity);
 static void amd_sched_wakeup(struct amd_gpu_scheduler *sched);
+static void amd_sched_process_job(struct fence *f, struct fence_cb *cb);
 
 struct kmem_cache *sched_fence_slab;
 atomic_t sched_fence_slab_ref = ATOMIC_INIT(0);
@@ -140,7 +141,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 		return r;
 
 	atomic_set(&entity->fence_seq, 0);
-	entity->fence_context = fence_context_alloc(1);
+	entity->fence_context = fence_context_alloc(2);
 
 	return 0;
 }
@@ -251,17 +252,21 @@ static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
 
 	s_fence = to_amd_sched_fence(fence);
 	if (s_fence && s_fence->sched == sched) {
-		/* Fence is from the same scheduler */
-		if (test_bit(AMD_SCHED_FENCE_SCHEDULED_BIT, &fence->flags)) {
-			/* Ignore it when it is already scheduled */
-			fence_put(entity->dependency);
-			return false;
-		}
 
-		/* Wait for fence to be scheduled */
-		entity->cb.func = amd_sched_entity_clear_dep;
-		list_add_tail(&entity->cb.node, &s_fence->scheduled_cb);
-		return true;
+		/*
+		 * Fence is from the same scheduler, only need to wait for
+		 * it to be scheduled
+		 */
+		fence = fence_get(&s_fence->scheduled);
+		fence_put(entity->dependency);
+		entity->dependency = fence;
+		if (!fence_add_callback(fence, &entity->cb,
+					amd_sched_entity_clear_dep))
+			return true;
+
+		/* Ignore it when it is already scheduled */
+		fence_put(fence);
+		return false;
 	}
 
 	if (!fence_add_callback(entity->dependency, &entity->cb,
@@ -319,46 +324,114 @@ static bool amd_sched_entity_in(struct amd_sched_job *sched_job)
 	return added;
 }
 
-static void amd_sched_free_job(struct fence *f, struct fence_cb *cb) {
-	struct amd_sched_job *job = container_of(cb, struct amd_sched_job, cb_free_job);
-	schedule_work(&job->work_free_job);
-}
-
 /* job_finish is called after hw fence signaled, and
  * the job had already been deleted from ring_mirror_list
  */
-void amd_sched_job_finish(struct amd_sched_job *s_job)
+static void amd_sched_job_finish(struct work_struct *work)
 {
-	struct amd_sched_job *next;
+	struct amd_sched_job *s_job = container_of(work, struct amd_sched_job,
+						   finish_work);
 	struct amd_gpu_scheduler *sched = s_job->sched;
 
+	/* remove job from ring_mirror_list */
+	spin_lock(&sched->job_list_lock);
+	list_del_init(&s_job->node);
 	if (sched->timeout != MAX_SCHEDULE_TIMEOUT) {
-		if (cancel_delayed_work(&s_job->work_tdr))
-			amd_sched_job_put(s_job);
+		struct amd_sched_job *next;
+
+		spin_unlock(&sched->job_list_lock);
+		cancel_delayed_work_sync(&s_job->work_tdr);
+		spin_lock(&sched->job_list_lock);
 
 		/* queue TDR for next job */
 		next = list_first_entry_or_null(&sched->ring_mirror_list,
 						struct amd_sched_job, node);
 
-		if (next) {
-			INIT_DELAYED_WORK(&next->work_tdr, s_job->timeout_callback);
-			amd_sched_job_get(next);
+		if (next)
 			schedule_delayed_work(&next->work_tdr, sched->timeout);
-		}
 	}
+	spin_unlock(&sched->job_list_lock);
+	sched->ops->free_job(s_job);
 }
 
-void amd_sched_job_begin(struct amd_sched_job *s_job)
+static void amd_sched_job_finish_cb(struct fence *f, struct fence_cb *cb)
+{
+	struct amd_sched_job *job = container_of(cb, struct amd_sched_job,
+						 finish_cb);
+	schedule_work(&job->finish_work);
+}
+
+static void amd_sched_job_begin(struct amd_sched_job *s_job)
 {
 	struct amd_gpu_scheduler *sched = s_job->sched;
 
+	spin_lock(&sched->job_list_lock);
+	list_add_tail(&s_job->node, &sched->ring_mirror_list);
 	if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&
-		list_first_entry_or_null(&sched->ring_mirror_list, struct amd_sched_job, node) == s_job)
-	{
-		INIT_DELAYED_WORK(&s_job->work_tdr, s_job->timeout_callback);
-		amd_sched_job_get(s_job);
+	    list_first_entry_or_null(&sched->ring_mirror_list,
+				     struct amd_sched_job, node) == s_job)
+		schedule_delayed_work(&s_job->work_tdr, sched->timeout);
+	spin_unlock(&sched->job_list_lock);
+}
+
+static void amd_sched_job_timedout(struct work_struct *work)
+{
+	struct amd_sched_job *job = container_of(work, struct amd_sched_job,
+						 work_tdr.work);
+
+	job->sched->ops->timedout_job(job);
+}
+
+void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched)
+{
+	struct amd_sched_job *s_job;
+
+	spin_lock(&sched->job_list_lock);
+	list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
+		if (fence_remove_callback(s_job->s_fence->parent, &s_job->s_fence->cb)) {
+			fence_put(s_job->s_fence->parent);
+			s_job->s_fence->parent = NULL;
+		}
+	}
+	atomic_set(&sched->hw_rq_count, 0);
+	spin_unlock(&sched->job_list_lock);
+}
+
+void amd_sched_job_recovery(struct amd_gpu_scheduler *sched)
+{
+	struct amd_sched_job *s_job, *tmp;
+	int r;
+
+	spin_lock(&sched->job_list_lock);
+	s_job = list_first_entry_or_null(&sched->ring_mirror_list,
+					 struct amd_sched_job, node);
+	if (s_job)
 		schedule_delayed_work(&s_job->work_tdr, sched->timeout);
+
+	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
+		struct amd_sched_fence *s_fence = s_job->s_fence;
+		struct fence *fence;
+
+		spin_unlock(&sched->job_list_lock);
+		fence = sched->ops->run_job(s_job);
+		atomic_inc(&sched->hw_rq_count);
+		if (fence) {
+			s_fence->parent = fence_get(fence);
+			r = fence_add_callback(fence, &s_fence->cb,
+					       amd_sched_process_job);
+			if (r == -ENOENT)
+				amd_sched_process_job(fence, &s_fence->cb);
+			else if (r)
+				DRM_ERROR("fence add callback failed (%d)\n",
+					  r);
+			fence_put(fence);
+		} else {
+			DRM_ERROR("Failed to run job!\n");
+			amd_sched_process_job(NULL, &s_fence->cb);
+		}
+		spin_lock(&sched->job_list_lock);
 	}
+	spin_unlock(&sched->job_list_lock);
 }
 
 /**
@@ -372,36 +445,29 @@ void amd_sched_entity_push_job(struct amd_sched_job *sched_job)
 {
 	struct amd_sched_entity *entity = sched_job->s_entity;
 
-	sched_job->use_sched = 1;
-	fence_add_callback(&sched_job->s_fence->base,
-					&sched_job->cb_free_job, amd_sched_free_job);
 	trace_amd_sched_job(sched_job);
+	fence_add_callback(&sched_job->s_fence->finished, &sched_job->finish_cb,
+			   amd_sched_job_finish_cb);
 	wait_event(entity->sched->job_scheduled,
 		   amd_sched_entity_in(sched_job));
 }
 
 /* init a sched_job with basic field */
 int amd_sched_job_init(struct amd_sched_job *job,
-						struct amd_gpu_scheduler *sched,
-						struct amd_sched_entity *entity,
-						void (*timeout_cb)(struct work_struct *work),
-						void (*free_cb)(struct kref *refcount),
-						void *owner, struct fence **fence)
+		       struct amd_gpu_scheduler *sched,
+		       struct amd_sched_entity *entity,
+		       void *owner)
 {
-	INIT_LIST_HEAD(&job->node);
-	kref_init(&job->refcount);
 	job->sched = sched;
 	job->s_entity = entity;
 	job->s_fence = amd_sched_fence_create(entity, owner);
 	if (!job->s_fence)
 		return -ENOMEM;
 
-	job->s_fence->s_job = job;
-	job->timeout_callback = timeout_cb;
-	job->free_callback = free_cb;
+	INIT_WORK(&job->finish_work, amd_sched_job_finish);
+	INIT_LIST_HEAD(&job->node);
+	INIT_DELAYED_WORK(&job->work_tdr, amd_sched_job_timedout);
 
-	if (fence)
-		*fence = &job->s_fence->base;
 	return 0;
 }
 
@@ -450,23 +516,25 @@ static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 	struct amd_sched_fence *s_fence =
 		container_of(cb, struct amd_sched_fence, cb);
 	struct amd_gpu_scheduler *sched = s_fence->sched;
-	unsigned long flags;
 
 	atomic_dec(&sched->hw_rq_count);
-
-	/* remove job from ring_mirror_list */
-	spin_lock_irqsave(&sched->job_list_lock, flags);
-	list_del_init(&s_fence->s_job->node);
-	sched->ops->finish_job(s_fence->s_job);
-	spin_unlock_irqrestore(&sched->job_list_lock, flags);
-
-	amd_sched_fence_signal(s_fence);
+	amd_sched_fence_finished(s_fence);
 
 	trace_amd_sched_process_job(s_fence);
-	fence_put(&s_fence->base);
+	fence_put(&s_fence->finished);
 	wake_up_interruptible(&sched->wake_up_worker);
 }
 
+static bool amd_sched_blocked(struct amd_gpu_scheduler *sched)
+{
+	if (kthread_should_park()) {
+		kthread_parkme();
+		return true;
+	}
+
+	return false;
+}
+
 static int amd_sched_main(void *param)
 {
 	struct sched_param sparam = {.sched_priority = 1};
@@ -476,14 +544,15 @@ static int amd_sched_main(void *param)
 	sched_setscheduler(current, SCHED_FIFO, &sparam);
 
 	while (!kthread_should_stop()) {
-		struct amd_sched_entity *entity;
+		struct amd_sched_entity *entity = NULL;
 		struct amd_sched_fence *s_fence;
 		struct amd_sched_job *sched_job;
 		struct fence *fence;
 
 		wait_event_interruptible(sched->wake_up_worker,
-			(entity = amd_sched_select_entity(sched)) ||
-			kthread_should_stop());
+					 (!amd_sched_blocked(sched) &&
+					  (entity = amd_sched_select_entity(sched))) ||
+					 kthread_should_stop());
 
 		if (!entity)
 			continue;
@@ -495,16 +564,19 @@ static int amd_sched_main(void *param)
 		s_fence = sched_job->s_fence;
 
 		atomic_inc(&sched->hw_rq_count);
-		amd_sched_job_pre_schedule(sched, sched_job);
+		amd_sched_job_begin(sched_job);
+
 		fence = sched->ops->run_job(sched_job);
 		amd_sched_fence_scheduled(s_fence);
 		if (fence) {
+			s_fence->parent = fence_get(fence);
 			r = fence_add_callback(fence, &s_fence->cb,
 					       amd_sched_process_job);
 			if (r == -ENOENT)
 				amd_sched_process_job(fence, &s_fence->cb);
 			else if (r)
-				DRM_ERROR("fence add callback failed (%d)\n", r);
+				DRM_ERROR("fence add callback failed (%d)\n",
+					  r);
 			fence_put(fence);
 		} else {
 			DRM_ERROR("Failed to run job!\n");
diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
index 070095a9433c..7cbbbfb502ef 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_SCHED_FENCE_SCHEDULED_BIT	FENCE_FLAG_USER_BITS
-
 struct amd_gpu_scheduler;
 struct amd_sched_rq;
 
@@ -68,36 +66,34 @@ struct amd_sched_rq {
 };
 
 struct amd_sched_fence {
-	struct fence                    base;
+	struct fence                    scheduled;
+	struct fence                    finished;
 	struct fence_cb                 cb;
-	struct list_head		scheduled_cb;
+	struct fence                    *parent;
 	struct amd_gpu_scheduler	*sched;
 	spinlock_t			lock;
 	void                            *owner;
-	struct amd_sched_job	*s_job;
 };
 
 struct amd_sched_job {
-	struct kref refcount;
 	struct amd_gpu_scheduler        *sched;
 	struct amd_sched_entity         *s_entity;
 	struct amd_sched_fence          *s_fence;
-	bool	use_sched;	/* true if the job goes to scheduler */
-	struct fence_cb                cb_free_job;
-	struct work_struct             work_free_job;
-	struct list_head			   node;
-	struct delayed_work work_tdr;
-	void (*timeout_callback) (struct work_struct *work);
-	void (*free_callback)(struct kref *refcount);
+	struct fence_cb			finish_cb;
+	struct work_struct		finish_work;
+	struct list_head		node;
+	struct delayed_work		work_tdr;
 };
 
-extern const struct fence_ops amd_sched_fence_ops;
+extern const struct fence_ops amd_sched_fence_ops_scheduled;
+extern const struct fence_ops amd_sched_fence_ops_finished;
 static inline struct amd_sched_fence *to_amd_sched_fence(struct fence *f)
 {
-	struct amd_sched_fence *__f = container_of(f, struct amd_sched_fence, base);
+	if (f->ops == &amd_sched_fence_ops_scheduled)
+		return container_of(f, struct amd_sched_fence, scheduled);
 
-	if (__f->base.ops == &amd_sched_fence_ops)
-		return __f;
+	if (f->ops == &amd_sched_fence_ops_finished)
+		return container_of(f, struct amd_sched_fence, finished);
 
 	return NULL;
 }
@@ -109,8 +105,8 @@ static inline struct amd_sched_fence *to_amd_sched_fence(struct fence *f)
 struct amd_sched_backend_ops {
 	struct fence *(*dependency)(struct amd_sched_job *sched_job);
 	struct fence *(*run_job)(struct amd_sched_job *sched_job);
-	void (*begin_job)(struct amd_sched_job *sched_job);
-	void (*finish_job)(struct amd_sched_job *sched_job);
+	void (*timedout_job)(struct amd_sched_job *sched_job);
+	void (*free_job)(struct amd_sched_job *sched_job);
 };
 
 enum amd_sched_priority {
@@ -152,25 +148,11 @@ void 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, void *owner);
 void amd_sched_fence_scheduled(struct amd_sched_fence *fence);
-void amd_sched_fence_signal(struct amd_sched_fence *fence);
+void amd_sched_fence_finished(struct amd_sched_fence *fence);
 int amd_sched_job_init(struct amd_sched_job *job,
-					struct amd_gpu_scheduler *sched,
-					struct amd_sched_entity *entity,
-					void (*timeout_cb)(struct work_struct *work),
-					void (*free_cb)(struct kref* refcount),
-					void *owner, struct fence **fence);
-void amd_sched_job_pre_schedule(struct amd_gpu_scheduler *sched ,
-								struct amd_sched_job *s_job);
-void amd_sched_job_finish(struct amd_sched_job *s_job);
-void amd_sched_job_begin(struct amd_sched_job *s_job);
-static inline void amd_sched_job_get(struct amd_sched_job *job) {
-	if (job)
-		kref_get(&job->refcount);
-}
-
-static inline void amd_sched_job_put(struct amd_sched_job *job) {
-	if (job)
-		kref_put(&job->refcount, job->free_callback);
-}
-
+		       struct amd_gpu_scheduler *sched,
+		       struct amd_sched_entity *entity,
+		       void *owner);
+void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched);
+void amd_sched_job_recovery(struct amd_gpu_scheduler *sched);
 #endif
diff --git a/drivers/gpu/drm/amd/scheduler/sched_fence.c b/drivers/gpu/drm/amd/scheduler/sched_fence.c
index 2a732c490375..6b63beaf7574 100644
--- a/drivers/gpu/drm/amd/scheduler/sched_fence.c
+++ b/drivers/gpu/drm/amd/scheduler/sched_fence.c
@@ -27,7 +27,8 @@
 #include <drm/drmP.h>
 #include "gpu_scheduler.h"
 
-struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *s_entity, void *owner)
+struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *entity,
+					       void *owner)
 {
 	struct amd_sched_fence *fence = NULL;
 	unsigned seq;
@@ -36,46 +37,37 @@ struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *s_entity
 	if (fence == NULL)
 		return NULL;
 
-	INIT_LIST_HEAD(&fence->scheduled_cb);
 	fence->owner = owner;
-	fence->sched = s_entity->sched;
+	fence->sched = entity->sched;
 	spin_lock_init(&fence->lock);
 
-	seq = atomic_inc_return(&s_entity->fence_seq);
-	fence_init(&fence->base, &amd_sched_fence_ops, &fence->lock,
-		   s_entity->fence_context, seq);
+	seq = atomic_inc_return(&entity->fence_seq);
+	fence_init(&fence->scheduled, &amd_sched_fence_ops_scheduled,
+		   &fence->lock, entity->fence_context, seq);
+	fence_init(&fence->finished, &amd_sched_fence_ops_finished,
+		   &fence->lock, entity->fence_context + 1, seq);
 
 	return fence;
 }
 
-void amd_sched_fence_signal(struct amd_sched_fence *fence)
+void amd_sched_fence_scheduled(struct amd_sched_fence *fence)
 {
-	int ret = fence_signal(&fence->base);
+	int ret = fence_signal(&fence->scheduled);
+
 	if (!ret)
-		FENCE_TRACE(&fence->base, "signaled from irq context\n");
+		FENCE_TRACE(&fence->scheduled, "signaled from irq context\n");
 	else
-		FENCE_TRACE(&fence->base, "was already signaled\n");
-}
-
-void amd_sched_job_pre_schedule(struct amd_gpu_scheduler *sched ,
-				struct amd_sched_job *s_job)
-{
-	unsigned long flags;
-	spin_lock_irqsave(&sched->job_list_lock, flags);
-	list_add_tail(&s_job->node, &sched->ring_mirror_list);
-	sched->ops->begin_job(s_job);
-	spin_unlock_irqrestore(&sched->job_list_lock, flags);
+		FENCE_TRACE(&fence->scheduled, "was already signaled\n");
 }
 
-void amd_sched_fence_scheduled(struct amd_sched_fence *s_fence)
+void amd_sched_fence_finished(struct amd_sched_fence *fence)
 {
-	struct fence_cb *cur, *tmp;
+	int ret = fence_signal(&fence->finished);
 
-	set_bit(AMD_SCHED_FENCE_SCHEDULED_BIT, &s_fence->base.flags);
-	list_for_each_entry_safe(cur, tmp, &s_fence->scheduled_cb, node) {
-		list_del_init(&cur->node);
-		cur->func(&s_fence->base, cur);
-	}
+	if (!ret)
+		FENCE_TRACE(&fence->finished, "signaled from irq context\n");
+	else
+		FENCE_TRACE(&fence->finished, "was already signaled\n");
 }
 
 static const char *amd_sched_fence_get_driver_name(struct fence *fence)
@@ -105,6 +97,8 @@ static void amd_sched_fence_free(struct rcu_head *rcu)
 {
 	struct fence *f = container_of(rcu, struct fence, rcu);
 	struct amd_sched_fence *fence = to_amd_sched_fence(f);
+
+	fence_put(fence->parent);
 	kmem_cache_free(sched_fence_slab, fence);
 }
 
@@ -116,16 +110,41 @@ static void amd_sched_fence_free(struct rcu_head *rcu)
  * This function is called when the reference count becomes zero.
  * It just RCU schedules freeing up the fence.
  */
-static void amd_sched_fence_release(struct fence *f)
+static void amd_sched_fence_release_scheduled(struct fence *f)
 {
-	call_rcu(&f->rcu, amd_sched_fence_free);
+	struct amd_sched_fence *fence = to_amd_sched_fence(f);
+
+	call_rcu(&fence->finished.rcu, amd_sched_fence_free);
 }
 
-const struct fence_ops amd_sched_fence_ops = {
+/**
+ * amd_sched_fence_release_scheduled - drop extra reference
+ *
+ * @f: fence
+ *
+ * Drop the extra reference from the scheduled fence to the base fence.
+ */
+static void amd_sched_fence_release_finished(struct fence *f)
+{
+	struct amd_sched_fence *fence = to_amd_sched_fence(f);
+
+	fence_put(&fence->scheduled);
+}
+
+const struct fence_ops amd_sched_fence_ops_scheduled = {
+	.get_driver_name = amd_sched_fence_get_driver_name,
+	.get_timeline_name = amd_sched_fence_get_timeline_name,
+	.enable_signaling = amd_sched_fence_enable_signaling,
+	.signaled = NULL,
+	.wait = fence_default_wait,
+	.release = amd_sched_fence_release_scheduled,
+};
+
+const struct fence_ops amd_sched_fence_ops_finished = {
 	.get_driver_name = amd_sched_fence_get_driver_name,
 	.get_timeline_name = amd_sched_fence_get_timeline_name,
 	.enable_signaling = amd_sched_fence_enable_signaling,
 	.signaled = NULL,
 	.wait = fence_default_wait,
-	.release = amd_sched_fence_release,
+	.release = amd_sched_fence_release_finished,
 };