diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index 44da025502acad3dbea7b156eeb10e86927e0977..567597469a8a30adc41a477832cfc15c3673079f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -684,6 +684,8 @@ int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev,
 		goto err_ib_sched;
 	}
 
+	/* Drop the initial kref_init count (see drm_sched_main as example) */
+	dma_fence_put(f);
 	ret = dma_fence_wait(f, false);
 
 err_ib_sched:
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 9d2395ab4f70bc2b5f2d6bb3289b7990a49912e8..0a1c47ccd1ea72c1e0cf1cefd8b80c70aab1c592 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -5010,16 +5010,32 @@ static void amdgpu_device_recheck_guilty_jobs(
 
 		/* clear job's guilty and depend the folowing step to decide the real one */
 		drm_sched_reset_karma(s_job);
-		/* for the real bad job, it will be resubmitted twice, adding a dma_fence_get
-		 * to make sure fence is balanced */
-		dma_fence_get(s_job->s_fence->parent);
 		drm_sched_resubmit_jobs_ext(&ring->sched, 1);
 
+		if (!s_job->s_fence->parent) {
+			DRM_WARN("Failed to get a HW fence for job!");
+			continue;
+		}
+
 		ret = dma_fence_wait_timeout(s_job->s_fence->parent, false, ring->sched.timeout);
 		if (ret == 0) { /* timeout */
 			DRM_ERROR("Found the real bad job! ring:%s, job_id:%llx\n",
 						ring->sched.name, s_job->id);
 
+
+			amdgpu_fence_driver_isr_toggle(adev, true);
+
+			/* Clear this failed job from fence array */
+			amdgpu_fence_driver_clear_job_fences(ring);
+
+			amdgpu_fence_driver_isr_toggle(adev, false);
+
+			/* Since the job won't signal and we go for
+			 * another resubmit drop this parent pointer
+			 */
+			dma_fence_put(s_job->s_fence->parent);
+			s_job->s_fence->parent = NULL;
+
 			/* set guilty */
 			drm_sched_increase_karma(s_job);
 retry:
@@ -5048,7 +5064,6 @@ static void amdgpu_device_recheck_guilty_jobs(
 
 		/* got the hw fence, signal finished fence */
 		atomic_dec(ring->sched.score);
-		dma_fence_put(s_job->s_fence->parent);
 		dma_fence_get(&s_job->s_fence->finished);
 		dma_fence_signal(&s_job->s_fence->finished);
 		dma_fence_put(&s_job->s_fence->finished);
@@ -5221,8 +5236,8 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
 	 *
 	 * job->base holds a reference to parent fence
 	 */
-	if (job && job->base.s_fence->parent &&
-	    dma_fence_is_signaled(job->base.s_fence->parent)) {
+	if (job && (job->hw_fence.ops != NULL) &&
+	    dma_fence_is_signaled(&job->hw_fence)) {
 		job_signaled = true;
 		dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
 		goto skip_hw_reset;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index c1d04ea3c67f38625896ead549d3dd48905aa1e2..39597ab807d105bdff0dd1d0e9aa181ea10108f0 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -164,11 +164,16 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd
 	if (job && job->job_run_counter) {
 		/* reinit seq for resubmitted jobs */
 		fence->seqno = seq;
+		/* TO be inline with external fence creation and other drivers */
+		dma_fence_get(fence);
 	} else {
-		if (job)
+		if (job) {
 			dma_fence_init(fence, &amdgpu_job_fence_ops,
 				       &ring->fence_drv.lock,
 				       adev->fence_context + ring->idx, seq);
+			/* Against remove in amdgpu_job_{free, free_cb} */
+			dma_fence_get(fence);
+		}
 		else
 			dma_fence_init(fence, &amdgpu_fence_ops,
 				       &ring->fence_drv.lock,
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 26ede765eed89ad90cf485aee991a7d4d3e423a3..22735790fe5020bf3792620c7433e0f4a2ed821e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -262,10 +262,6 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
 			DRM_ERROR("Error scheduling IBs (%d)\n", r);
 	}
 
-	if (!job->job_run_counter)
-		dma_fence_get(fence);
-	else if (finished->error < 0)
-		dma_fence_put(&job->hw_fence);
 	job->job_run_counter++;
 	amdgpu_job_free_resources(job);