1 files changed, 443 insertions, 4 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 44d87943e40a..bc9e81293165 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -952,6 +952,92 @@ out_unlock:
 	return retval;
 }
 
+/* suspend_single_queue does not lock the dqm like the
+ * evict_process_queues_cpsch or evict_process_queues_nocpsch. You should
+ * lock the dqm before calling, and unlock after calling.
+ *
+ * The reason we don't lock the dqm is because this function may be
+ * called on multiple queues in a loop, so rather than locking/unlocking
+ * multiple times, we will just keep the dqm locked for all of the calls.
+ */
+static int suspend_single_queue(struct device_queue_manager *dqm,
+				      struct kfd_process_device *pdd,
+				      struct queue *q)
+{
+	bool is_new;
+
+	if (q->properties.is_suspended)
+		return 0;
+
+	pr_debug("Suspending PASID %u queue [%i]\n",
+			pdd->process->pasid,
+			q->properties.queue_id);
+
+	is_new = q->properties.exception_status & KFD_EC_MASK(EC_QUEUE_NEW);
+
+	if (is_new || q->properties.is_being_destroyed) {
+		pr_debug("Suspend: skip %s queue id %i\n",
+				is_new ? "new" : "destroyed",
+				q->properties.queue_id);
+		return -EBUSY;
+	}
+
+	q->properties.is_suspended = true;
+	if (q->properties.is_active) {
+		if (dqm->dev->kfd->shared_resources.enable_mes) {
+			int r = remove_queue_mes(dqm, q, &pdd->qpd);
+
+			if (r)
+				return r;
+		}
+
+		decrement_queue_count(dqm, &pdd->qpd, q);
+		q->properties.is_active = false;
+	}
+
+	return 0;
+}
+
+/* resume_single_queue does not lock the dqm like the functions
+ * restore_process_queues_cpsch or restore_process_queues_nocpsch. You should
+ * lock the dqm before calling, and unlock after calling.
+ *
+ * The reason we don't lock the dqm is because this function may be
+ * called on multiple queues in a loop, so rather than locking/unlocking
+ * multiple times, we will just keep the dqm locked for all of the calls.
+ */
+static int resume_single_queue(struct device_queue_manager *dqm,
+				      struct qcm_process_device *qpd,
+				      struct queue *q)
+{
+	struct kfd_process_device *pdd;
+
+	if (!q->properties.is_suspended)
+		return 0;
+
+	pdd = qpd_to_pdd(qpd);
+
+	pr_debug("Restoring from suspend PASID %u queue [%i]\n",
+			    pdd->process->pasid,
+			    q->properties.queue_id);
+
+	q->properties.is_suspended = false;
+
+	if (QUEUE_IS_ACTIVE(q->properties)) {
+		if (dqm->dev->kfd->shared_resources.enable_mes) {
+			int r = add_queue_mes(dqm, q, &pdd->qpd);
+
+			if (r)
+				return r;
+		}
+
+		q->properties.is_active = true;
+		increment_queue_count(dqm, qpd, q);
+	}
+
+	return 0;
+}
+
 static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
 					struct qcm_process_device *qpd)
 {
@@ -1926,6 +2012,31 @@ static int execute_queues_cpsch(struct device_queue_manager *dqm,
 	return map_queues_cpsch(dqm);
 }
 
+static int wait_on_destroy_queue(struct device_queue_manager *dqm,
+				 struct queue *q)
+{
+	struct kfd_process_device *pdd = kfd_get_process_device_data(q->device,
+								q->process);
+	int ret = 0;
+
+	if (pdd->qpd.is_debug)
+		return ret;
+
+	q->properties.is_being_destroyed = true;
+
+	if (pdd->process->debug_trap_enabled && q->properties.is_suspended) {
+		dqm_unlock(dqm);
+		mutex_unlock(&q->process->mutex);
+		ret = wait_event_interruptible(dqm->destroy_wait,
+						!q->properties.is_suspended);
+
+		mutex_lock(&q->process->mutex);
+		dqm_lock(dqm);
+	}
+
+	return ret;
+}
+
 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
 				struct qcm_process_device *qpd,
 				struct queue *q)
@@ -1945,11 +2056,16 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
 				q->properties.queue_id);
 	}
 
-	retval = 0;
-
 	/* remove queue from list to prevent rescheduling after preemption */
 	dqm_lock(dqm);
 
+	retval = wait_on_destroy_queue(dqm, q);
+
+	if (retval) {
+		dqm_unlock(dqm);
+		return retval;
+	}
+
 	if (qpd->is_debug) {
 		/*
 		 * error, currently we do not allow to destroy a queue
@@ -1996,7 +2112,14 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
 
 	dqm_unlock(dqm);
 
-	/* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */
+	/*
+	 * Do free_mqd and raise delete event after dqm_unlock(dqm) to avoid
+	 * circular locking
+	 */
+	kfd_dbg_ev_raise(KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE),
+				qpd->pqm->process, q->device,
+				-1, false, NULL, 0);
+
 	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
 
 	return retval;
@@ -2461,8 +2584,10 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev)
 		goto out_free;
 	}
 
-	if (!dqm->ops.initialize(dqm))
+	if (!dqm->ops.initialize(dqm)) {
+		init_waitqueue_head(&dqm->destroy_wait);
 		return dqm;
+	}
 
 out_free:
 	kfree(dqm);
@@ -2602,6 +2727,320 @@ out_unlock:
 	return r;
 }
 
+#define QUEUE_NOT_FOUND		-1
+/* invalidate queue operation in array */
+static void q_array_invalidate(uint32_t num_queues, uint32_t *queue_ids)
+{
+	int i;
+
+	for (i = 0; i < num_queues; i++)
+		queue_ids[i] |= KFD_DBG_QUEUE_INVALID_MASK;
+}
+
+/* find queue index in array */
+static int q_array_get_index(unsigned int queue_id,
+		uint32_t num_queues,
+		uint32_t *queue_ids)
+{
+	int i;
+
+	for (i = 0; i < num_queues; i++)
+		if (queue_id == (queue_ids[i] & ~KFD_DBG_QUEUE_INVALID_MASK))
+			return i;
+
+	return QUEUE_NOT_FOUND;
+}
+
+struct copy_context_work_handler_workarea {
+	struct work_struct copy_context_work;
+	struct kfd_process *p;
+};
+
+static void copy_context_work_handler (struct work_struct *work)
+{
+	struct copy_context_work_handler_workarea *workarea;
+	struct mqd_manager *mqd_mgr;
+	struct queue *q;
+	struct mm_struct *mm;
+	struct kfd_process *p;
+	uint32_t tmp_ctl_stack_used_size, tmp_save_area_used_size;
+	int i;
+
+	workarea = container_of(work,
+			struct copy_context_work_handler_workarea,
+			copy_context_work);
+
+	p = workarea->p;
+	mm = get_task_mm(p->lead_thread);
+
+	if (!mm)
+		return;
+
+	kthread_use_mm(mm);
+	for (i = 0; i < p->n_pdds; i++) {
+		struct kfd_process_device *pdd = p->pdds[i];
+		struct device_queue_manager *dqm = pdd->dev->dqm;
+		struct qcm_process_device *qpd = &pdd->qpd;
+
+		list_for_each_entry(q, &qpd->queues_list, list) {
+			mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP];
+
+			/* We ignore the return value from get_wave_state
+			 * because
+			 * i) right now, it always returns 0, and
+			 * ii) if we hit an error, we would continue to the
+			 *      next queue anyway.
+			 */
+			mqd_mgr->get_wave_state(mqd_mgr,
+					q->mqd,
+					&q->properties,
+					(void __user *)	q->properties.ctx_save_restore_area_address,
+					&tmp_ctl_stack_used_size,
+					&tmp_save_area_used_size);
+		}
+	}
+	kthread_unuse_mm(mm);
+	mmput(mm);
+}
+
+static uint32_t *get_queue_ids(uint32_t num_queues, uint32_t *usr_queue_id_array)
+{
+	size_t array_size = num_queues * sizeof(uint32_t);
+	uint32_t *queue_ids = NULL;
+
+	if (!usr_queue_id_array)
+		return NULL;
+
+	queue_ids = kzalloc(array_size, GFP_KERNEL);
+	if (!queue_ids)
+		return ERR_PTR(-ENOMEM);
+
+	if (copy_from_user(queue_ids, usr_queue_id_array, array_size))
+		return ERR_PTR(-EFAULT);
+
+	return queue_ids;
+}
+
+int resume_queues(struct kfd_process *p,
+		uint32_t num_queues,
+		uint32_t *usr_queue_id_array)
+{
+	uint32_t *queue_ids = NULL;
+	int total_resumed = 0;
+	int i;
+
+	if (usr_queue_id_array) {
+		queue_ids = get_queue_ids(num_queues, usr_queue_id_array);
+
+		if (IS_ERR(queue_ids))
+			return PTR_ERR(queue_ids);
+
+		/* mask all queues as invalid.  unmask per successful request */
+		q_array_invalidate(num_queues, queue_ids);
+	}
+
+	for (i = 0; i < p->n_pdds; i++) {
+		struct kfd_process_device *pdd = p->pdds[i];
+		struct device_queue_manager *dqm = pdd->dev->dqm;
+		struct qcm_process_device *qpd = &pdd->qpd;
+		struct queue *q;
+		int r, per_device_resumed = 0;
+
+		dqm_lock(dqm);
+
+		/* unmask queues that resume or already resumed as valid */
+		list_for_each_entry(q, &qpd->queues_list, list) {
+			int q_idx = QUEUE_NOT_FOUND;
+
+			if (queue_ids)
+				q_idx = q_array_get_index(
+						q->properties.queue_id,
+						num_queues,
+						queue_ids);
+
+			if (!queue_ids || q_idx != QUEUE_NOT_FOUND) {
+				int err = resume_single_queue(dqm, &pdd->qpd, q);
+
+				if (queue_ids) {
+					if (!err) {
+						queue_ids[q_idx] &=
+							~KFD_DBG_QUEUE_INVALID_MASK;
+					} else {
+						queue_ids[q_idx] |=
+							KFD_DBG_QUEUE_ERROR_MASK;
+						break;
+					}
+				}
+
+				if (dqm->dev->kfd->shared_resources.enable_mes) {
+					wake_up_all(&dqm->destroy_wait);
+					if (!err)
+						total_resumed++;
+				} else {
+					per_device_resumed++;
+				}
+			}
+		}
+
+		if (!per_device_resumed) {
+			dqm_unlock(dqm);
+			continue;
+		}
+
+		r = execute_queues_cpsch(dqm,
+					KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
+					0,
+					USE_DEFAULT_GRACE_PERIOD);
+		if (r) {
+			pr_err("Failed to resume process queues\n");
+			if (queue_ids) {
+				list_for_each_entry(q, &qpd->queues_list, list) {
+					int q_idx = q_array_get_index(
+							q->properties.queue_id,
+							num_queues,
+							queue_ids);
+
+					/* mask queue as error on resume fail */
+					if (q_idx != QUEUE_NOT_FOUND)
+						queue_ids[q_idx] |=
+							KFD_DBG_QUEUE_ERROR_MASK;
+				}
+			}
+		} else {
+			wake_up_all(&dqm->destroy_wait);
+			total_resumed += per_device_resumed;
+		}
+
+		dqm_unlock(dqm);
+	}
+
+	if (queue_ids) {
+		if (copy_to_user((void __user *)usr_queue_id_array, queue_ids,
+				num_queues * sizeof(uint32_t)))
+			pr_err("copy_to_user failed on queue resume\n");
+
+		kfree(queue_ids);
+	}
+
+	return total_resumed;
+}
+
+int suspend_queues(struct kfd_process *p,
+			uint32_t num_queues,
+			uint32_t grace_period,
+			uint64_t exception_clear_mask,
+			uint32_t *usr_queue_id_array)
+{
+	uint32_t *queue_ids = get_queue_ids(num_queues, usr_queue_id_array);
+	int total_suspended = 0;
+	int i;
+
+	if (IS_ERR(queue_ids))
+		return PTR_ERR(queue_ids);
+
+	/* mask all queues as invalid.  umask on successful request */
+	q_array_invalidate(num_queues, queue_ids);
+
+	for (i = 0; i < p->n_pdds; i++) {
+		struct kfd_process_device *pdd = p->pdds[i];
+		struct device_queue_manager *dqm = pdd->dev->dqm;
+		struct qcm_process_device *qpd = &pdd->qpd;
+		struct queue *q;
+		int r, per_device_suspended = 0;
+
+		mutex_lock(&p->event_mutex);
+		dqm_lock(dqm);
+
+		/* unmask queues that suspend or already suspended */
+		list_for_each_entry(q, &qpd->queues_list, list) {
+			int q_idx = q_array_get_index(q->properties.queue_id,
+							num_queues,
+							queue_ids);
+
+			if (q_idx != QUEUE_NOT_FOUND) {
+				int err = suspend_single_queue(dqm, pdd, q);
+				bool is_mes = dqm->dev->kfd->shared_resources.enable_mes;
+
+				if (!err) {
+					queue_ids[q_idx] &= ~KFD_DBG_QUEUE_INVALID_MASK;
+					if (exception_clear_mask && is_mes)
+						q->properties.exception_status &=
+							~exception_clear_mask;
+
+					if (is_mes)
+						total_suspended++;
+					else
+						per_device_suspended++;
+				} else if (err != -EBUSY) {
+					r = err;
+					queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK;
+					break;
+				}
+			}
+		}
+
+		if (!per_device_suspended) {
+			dqm_unlock(dqm);
+			mutex_unlock(&p->event_mutex);
+			if (total_suspended)
+				amdgpu_amdkfd_debug_mem_fence(dqm->dev->adev);
+			continue;
+		}
+
+		r = execute_queues_cpsch(dqm,
+			KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0,
+			grace_period);
+
+		if (r)
+			pr_err("Failed to suspend process queues.\n");
+		else
+			total_suspended += per_device_suspended;
+
+		list_for_each_entry(q, &qpd->queues_list, list) {
+			int q_idx = q_array_get_index(q->properties.queue_id,
+						num_queues, queue_ids);
+
+			if (q_idx == QUEUE_NOT_FOUND)
+				continue;
+
+			/* mask queue as error on suspend fail */
+			if (r)
+				queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK;
+			else if (exception_clear_mask)
+				q->properties.exception_status &=
+							~exception_clear_mask;
+		}
+
+		dqm_unlock(dqm);
+		mutex_unlock(&p->event_mutex);
+		amdgpu_device_flush_hdp(dqm->dev->adev, NULL);
+	}
+
+	if (total_suspended) {
+		struct copy_context_work_handler_workarea copy_context_worker;
+
+		INIT_WORK_ONSTACK(
+				&copy_context_worker.copy_context_work,
+				copy_context_work_handler);
+
+		copy_context_worker.p = p;
+
+		schedule_work(&copy_context_worker.copy_context_work);
+
+
+		flush_work(&copy_context_worker.copy_context_work);
+		destroy_work_on_stack(&copy_context_worker.copy_context_work);
+	}
+
+	if (copy_to_user((void __user *)usr_queue_id_array, queue_ids,
+			num_queues * sizeof(uint32_t)))
+		pr_err("copy_to_user failed on queue suspend\n");
+
+	kfree(queue_ids);
+
+	return total_suspended;
+}
+
 int debug_lock_and_unmap(struct device_queue_manager *dqm)
 {
 	int r;