1 files changed, 29 insertions, 11 deletions

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
index f7f6df40875e..d7eb6ac37f62 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -930,7 +930,6 @@ static void uninitialize(struct device_queue_manager *dqm)
 	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
 		kfree(dqm->mqd_mgrs[i]);
 	mutex_destroy(&dqm->lock_hidden);
-	kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
 }
 
 static int start_nocpsch(struct device_queue_manager *dqm)
@@ -947,12 +946,19 @@ static int start_nocpsch(struct device_queue_manager *dqm)
 static int stop_nocpsch(struct device_queue_manager *dqm)
 {
 	if (dqm->dev->device_info->asic_family == CHIP_HAWAII)
-		pm_uninit(&dqm->packets);
+		pm_uninit(&dqm->packets, false);
 	dqm->sched_running = false;
 
 	return 0;
 }
 
+static void pre_reset(struct device_queue_manager *dqm)
+{
+	dqm_lock(dqm);
+	dqm->is_resetting = true;
+	dqm_unlock(dqm);
+}
+
 static int allocate_sdma_queue(struct device_queue_manager *dqm,
 				struct queue *q)
 {
@@ -1100,6 +1106,7 @@ static int start_cpsch(struct device_queue_manager *dqm)
 	dqm_lock(dqm);
 	/* clear hang status when driver try to start the hw scheduler */
 	dqm->is_hws_hang = false;
+	dqm->is_resetting = false;
 	dqm->sched_running = true;
 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
 	dqm_unlock(dqm);
@@ -1107,20 +1114,24 @@ static int start_cpsch(struct device_queue_manager *dqm)
 	return 0;
 fail_allocate_vidmem:
 fail_set_sched_resources:
-	pm_uninit(&dqm->packets);
+	pm_uninit(&dqm->packets, false);
 fail_packet_manager_init:
 	return retval;
 }
 
 static int stop_cpsch(struct device_queue_manager *dqm)
 {
+	bool hanging;
+
 	dqm_lock(dqm);
-	unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
+	if (!dqm->is_hws_hang)
+		unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
+	hanging = dqm->is_hws_hang || dqm->is_resetting;
 	dqm->sched_running = false;
 	dqm_unlock(dqm);
 
 	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
-	pm_uninit(&dqm->packets);
+	pm_uninit(&dqm->packets, hanging);
 
 	return 0;
 }
@@ -1352,8 +1363,17 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm,
 	/* should be timed out */
 	retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
 				queue_preemption_timeout_ms);
-	if (retval)
+	if (retval) {
+		pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
+		dqm->is_hws_hang = true;
+		/* It's possible we're detecting a HWS hang in the
+		 * middle of a GPU reset. No need to schedule another
+		 * reset in this case.
+		 */
+		if (!dqm->is_resetting)
+			schedule_work(&dqm->hw_exception_work);
 		return retval;
+	}
 
 	pm_release_ib(&dqm->packets);
 	dqm->active_runlist = false;
@@ -1371,12 +1391,8 @@ static int execute_queues_cpsch(struct device_queue_manager *dqm,
 	if (dqm->is_hws_hang)
 		return -EIO;
 	retval = unmap_queues_cpsch(dqm, filter, filter_param);
-	if (retval) {
-		pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
-		dqm->is_hws_hang = true;
-		schedule_work(&dqm->hw_exception_work);
+	if (retval)
 		return retval;
-	}
 
 	return map_queues_cpsch(dqm);
 }
@@ -1770,6 +1786,7 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
 		dqm->ops.initialize = initialize_cpsch;
 		dqm->ops.start = start_cpsch;
 		dqm->ops.stop = stop_cpsch;
+		dqm->ops.pre_reset = pre_reset;
 		dqm->ops.destroy_queue = destroy_queue_cpsch;
 		dqm->ops.update_queue = update_queue;
 		dqm->ops.register_process = register_process;
@@ -1788,6 +1805,7 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
 		/* initialize dqm for no cp scheduling */
 		dqm->ops.start = start_nocpsch;
 		dqm->ops.stop = stop_nocpsch;
+		dqm->ops.pre_reset = pre_reset;
 		dqm->ops.create_queue = create_queue_nocpsch;
 		dqm->ops.destroy_queue = destroy_queue_nocpsch;
 		dqm->ops.update_queue = update_queue;