drm/msm: rework inactive-work

Re-arrange things a bit so that we can get work requested after a bo
fence passes, like pageflip, done before retiring bo's.  Without any
sort of bo cache in userspace, some games can trigger hundred's of
transient bo's, which can cause retire to take a long time (5-10ms).
Obviously we want a bo cache.. but this cleanup will make things a
bit easier for atomic as well and makes things a bit cleaner.

Signed-off-by: Rob Clark <robdclark@gmail.com>
Acked-by: David Brown <davidb@codeaurora.org>

drivers/gpu/drm/msm/mdp4/mdp4_crtc.c

@@ -51,7 +51,7 @@ struct mdp4_crtc {
 
 	/* if there is a pending flip, these will be non-null: */
 	struct drm_pending_vblank_event *event;
-	struct work_struct pageflip_work;
+	struct msm_fence_cb pageflip_cb;
 
 	/* the fb that we currently hold a scanout ref to: */
 	struct drm_framebuffer *fb;
@@ -132,10 +132,10 @@ static void crtc_flush(struct drm_crtc *crtc)
 	mdp4_write(mdp4_kms, REG_MDP4_OVERLAY_FLUSH, flush);
 }
 
-static void pageflip_worker(struct work_struct *work)
+static void pageflip_cb(struct msm_fence_cb *cb)
 {
 	struct mdp4_crtc *mdp4_crtc =
-		container_of(work, struct mdp4_crtc, pageflip_work);
+		container_of(cb, struct mdp4_crtc, pageflip_cb);
 	struct drm_crtc *crtc = &mdp4_crtc->base;
 
 	mdp4_plane_set_scanout(mdp4_crtc->plane, crtc->fb);
@@ -397,8 +397,7 @@ static int mdp4_crtc_page_flip(struct drm_crtc *crtc,
 	mdp4_crtc->event = event;
 	update_fb(crtc, true, new_fb);
 
-	return msm_gem_queue_inactive_work(obj,
-			&mdp4_crtc->pageflip_work);
+	return msm_gem_queue_inactive_cb(obj, &mdp4_crtc->pageflip_cb);
 }
 
 static int mdp4_crtc_set_property(struct drm_crtc *crtc,
@@ -702,7 +701,7 @@ struct drm_crtc *mdp4_crtc_init(struct drm_device *dev,
 	ret = drm_flip_work_init(&mdp4_crtc->unref_cursor_work, 64,
 			"unref cursor", unref_cursor_worker);
 
-	INIT_WORK(&mdp4_crtc->pageflip_work, pageflip_worker);
+	INIT_FENCE_CB(&mdp4_crtc->pageflip_cb, pageflip_cb);
 
 	drm_crtc_init(dev, crtc, &mdp4_crtc_funcs);
 	drm_crtc_helper_add(crtc, &mdp4_crtc_helper_funcs);

drivers/gpu/drm/msm/msm_drv.c

@@ -187,6 +187,7 @@ static int msm_load(struct drm_device *dev, unsigned long flags)
 	init_waitqueue_head(&priv->fence_event);
 
 	INIT_LIST_HEAD(&priv->inactive_list);
+	INIT_LIST_HEAD(&priv->fence_cbs);
 
 	drm_mode_config_init(dev);
 
@@ -539,15 +540,36 @@ int msm_wait_fence_interruptable(struct drm_device *dev, uint32_t fence,
 	return ret;
 }
 
-/* call under struct_mutex */
+/* called from workqueue */
 void msm_update_fence(struct drm_device *dev, uint32_t fence)
 {
 	struct msm_drm_private *priv = dev->dev_private;
 
-	if (fence > priv->completed_fence) {
-		priv->completed_fence = fence;
-		wake_up_all(&priv->fence_event);
+	mutex_lock(&dev->struct_mutex);
+	priv->completed_fence = max(fence, priv->completed_fence);
+
+	while (!list_empty(&priv->fence_cbs)) {
+		struct msm_fence_cb *cb;
+
+		cb = list_first_entry(&priv->fence_cbs,
+				struct msm_fence_cb, work.entry);
+
+		if (cb->fence > priv->completed_fence)
+			break;
+
+		list_del_init(&cb->work.entry);
+		queue_work(priv->wq, &cb->work);
 	}
+
+	mutex_unlock(&dev->struct_mutex);
+
+	wake_up_all(&priv->fence_event);
+}
+
+void __msm_fence_worker(struct work_struct *work)
+{
+	struct msm_fence_cb *cb = container_of(work, struct msm_fence_cb, work);
+	cb->func(cb);
 }
 
 /*

drivers/gpu/drm/msm/msm_drv.h

@@ -73,6 +73,9 @@ struct msm_drm_private {
 
 	struct workqueue_struct *wq;
 
+	/* callbacks deferred until bo is inactive: */
+	struct list_head fence_cbs;
+
 	/* registered IOMMU domains: */
 	unsigned int num_iommus;
 	struct iommu_domain *iommus[NUM_DOMAINS];
@@ -97,6 +100,20 @@ struct msm_format {
 	uint32_t pixel_format;
 };
 
+/* callback from wq once fence has passed: */
+struct msm_fence_cb {
+	struct work_struct work;
+	uint32_t fence;
+	void (*func)(struct msm_fence_cb *cb);
+};
+
+void __msm_fence_worker(struct work_struct *work);
+
+#define INIT_FENCE_CB(_cb, _func)  do {                     \
+		INIT_WORK(&(_cb)->work, __msm_fence_worker); \
+		(_cb)->func = _func;                         \
+	} while (0)
+
 /* As there are different display controller blocks depending on the
  * snapdragon version, the kms support is split out and the appropriate
  * implementation is loaded at runtime.  The kms module is responsible
@@ -160,8 +177,8 @@ int msm_gem_prime_pin(struct drm_gem_object *obj);
 void msm_gem_prime_unpin(struct drm_gem_object *obj);
 void *msm_gem_vaddr_locked(struct drm_gem_object *obj);
 void *msm_gem_vaddr(struct drm_gem_object *obj);
-int msm_gem_queue_inactive_work(struct drm_gem_object *obj,
-		struct work_struct *work);
+int msm_gem_queue_inactive_cb(struct drm_gem_object *obj,
+		struct msm_fence_cb *cb);
 void msm_gem_move_to_active(struct drm_gem_object *obj,
 		struct msm_gpu *gpu, bool write, uint32_t fence);
 void msm_gem_move_to_inactive(struct drm_gem_object *obj);

drivers/gpu/drm/msm/msm_gem.c

@@ -309,7 +309,17 @@ int msm_gem_get_iova_locked(struct drm_gem_object *obj, int id,
 
 int msm_gem_get_iova(struct drm_gem_object *obj, int id, uint32_t *iova)
 {
+	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	int ret;
+
+	/* this is safe right now because we don't unmap until the
+	 * bo is deleted:
+	 */
+	if (msm_obj->domain[id].iova) {
+		*iova = msm_obj->domain[id].iova;
+		return 0;
+	}
+
 	mutex_lock(&obj->dev->struct_mutex);
 	ret = msm_gem_get_iova_locked(obj, id, iova);
 	mutex_unlock(&obj->dev->struct_mutex);
@@ -379,21 +389,25 @@ void *msm_gem_vaddr(struct drm_gem_object *obj)
 	return ret;
 }
 
-int msm_gem_queue_inactive_work(struct drm_gem_object *obj,
-		struct work_struct *work)
+/* setup callback for when bo is no longer busy..
+ * TODO probably want to differentiate read vs write..
+ */
+int msm_gem_queue_inactive_cb(struct drm_gem_object *obj,
+		struct msm_fence_cb *cb)
 {
 	struct drm_device *dev = obj->dev;
 	struct msm_drm_private *priv = dev->dev_private;
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	int ret = 0;
 
 	mutex_lock(&dev->struct_mutex);
-	if (!list_empty(&work->entry)) {
+	if (!list_empty(&cb->work.entry)) {
 		ret = -EINVAL;
 	} else if (is_active(msm_obj)) {
-		list_add_tail(&work->entry, &msm_obj->inactive_work);
+		cb->fence = max(msm_obj->read_fence, msm_obj->write_fence);
+		list_add_tail(&cb->work.entry, &priv->fence_cbs);
 	} else {
-		queue_work(priv->wq, work);
+		queue_work(priv->wq, &cb->work);
 	}
 	mutex_unlock(&dev->struct_mutex);
 
@@ -426,16 +440,6 @@ void msm_gem_move_to_inactive(struct drm_gem_object *obj)
 	msm_obj->write_fence = 0;
 	list_del_init(&msm_obj->mm_list);
 	list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
-
-	while (!list_empty(&msm_obj->inactive_work)) {
-		struct work_struct *work;
-
-		work = list_first_entry(&msm_obj->inactive_work,
-				struct work_struct, entry);
-
-		list_del_init(&work->entry);
-		queue_work(priv->wq, work);
-	}
 }
 
 int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op,
@@ -604,7 +608,6 @@ static int msm_gem_new_impl(struct drm_device *dev,
 	reservation_object_init(msm_obj->resv);
 
 	INIT_LIST_HEAD(&msm_obj->submit_entry);
-	INIT_LIST_HEAD(&msm_obj->inactive_work);
 	list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
 
 	*obj = &msm_obj->base;

drivers/gpu/drm/msm/msm_gem.h

@@ -45,9 +45,6 @@ struct msm_gem_object {
 	 */
 	struct list_head submit_entry;
 
-	/* work defered until bo is inactive: */
-	struct list_head inactive_work;
-
 	struct page **pages;
 	struct sg_table *sgt;
 	void *vaddr;

drivers/gpu/drm/msm/msm_gpu.c

@@ -268,6 +268,8 @@ static void retire_worker(struct work_struct *work)
 	struct drm_device *dev = gpu->dev;
 	uint32_t fence = gpu->funcs->last_fence(gpu);
 
+	msm_update_fence(gpu->dev, fence);
+
 	mutex_lock(&dev->struct_mutex);
 
 	while (!list_empty(&gpu->active_list)) {
@@ -287,8 +289,6 @@ static void retire_worker(struct work_struct *work)
 		}
 	}
 
-	msm_update_fence(gpu->dev, fence);
-
 	mutex_unlock(&dev->struct_mutex);
 }