drm/xe: Use ordered wq for preempt fence waiting

Preempt fences can sleep waiting for an exec queue suspend operation to
complete. If the system_unbound_wq is used for waiting and the number of
waiters exceeds max_active this will result in other users of the
system_unbound_wq getting starved. Use a device private work queue for
preempt fences to avoid starvation of the system_unbound_wq.

Even though suspend operations can complete out-of-order, all suspend
operations within a VM need to complete before the preempt rebind worker
can start. With that, use a device private ordered wq for preempt fence
waiting.

v2:
 - Add comment about cleanup on failure (Matt R)
 - Update commit message (Lucas)

Fixes: dd08ebf ("drm/xe: Introduce a new DRM driver for Intel GPUs")
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240401221913.139672-2-matthew.brost@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
(cherry picked from commit 37c15c4)
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>

Author: mbrost05

drivers/gpu/drm/xe/xe_device.c

@@ -193,6 +193,9 @@ static void xe_device_destroy(struct drm_device *dev, void *dummy)
 {
 	struct xe_device *xe = to_xe_device(dev);
 
+	if (xe->preempt_fence_wq)
+		destroy_workqueue(xe->preempt_fence_wq);
+
 	if (xe->ordered_wq)
 		destroy_workqueue(xe->ordered_wq);
 
@@ -258,9 +261,15 @@ struct xe_device *xe_device_create(struct pci_dev *pdev,
 	INIT_LIST_HEAD(&xe->pinned.external_vram);
 	INIT_LIST_HEAD(&xe->pinned.evicted);
 
+	xe->preempt_fence_wq = alloc_ordered_workqueue("xe-preempt-fence-wq", 0);
 	xe->ordered_wq = alloc_ordered_workqueue("xe-ordered-wq", 0);
 	xe->unordered_wq = alloc_workqueue("xe-unordered-wq", 0, 0);
-	if (!xe->ordered_wq || !xe->unordered_wq) {
+	if (!xe->ordered_wq || !xe->unordered_wq ||
+	    !xe->preempt_fence_wq) {
+		/*
+		 * Cleanup done in xe_device_destroy via
+		 * drmm_add_action_or_reset register above
+		 */
 		drm_err(&xe->drm, "Failed to allocate xe workqueues\n");
 		err = -ENOMEM;
 		goto err;

drivers/gpu/drm/xe/xe_device_types.h

@@ -363,6 +363,9 @@ struct xe_device {
 	/** @ufence_wq: user fence wait queue */
 	wait_queue_head_t ufence_wq;
 
+	/** @preempt_fence_wq: used to serialize preempt fences */
+	struct workqueue_struct *preempt_fence_wq;
+
 	/** @ordered_wq: used to serialize compute mode resume */
 	struct workqueue_struct *ordered_wq;
 

drivers/gpu/drm/xe/xe_preempt_fence.c

@@ -49,7 +49,7 @@ static bool preempt_fence_enable_signaling(struct dma_fence *fence)
 	struct xe_exec_queue *q = pfence->q;
 
 	pfence->error = q->ops->suspend(q);
-	queue_work(system_unbound_wq, &pfence->preempt_work);
+	queue_work(q->vm->xe->preempt_fence_wq, &pfence->preempt_work);
 	return true;
 }