[PATCH 05/10] staging: drm/omap: defer unpin until scanout completes

[Rob Clark <rob.clark@linaro.org>]

From: Rob Clark <rob@ti.com>

When flipping, defer unpinning until scanout completes, as indicated
by the appropriate END_WIN irq.

This also re-organizes things a bit, in replacing omap_fb_{pin,unpin}
with omap_fb_replace(), to make it easier to add support for scanout
synchronized DMM refill mode (flipping by just reprogramming DMM
synchronized with DSS scanout).

Signed-off-by: Rob Clark <rob@ti.com>
---
 drivers/staging/omapdrm/omap_drv.h   |    5 +-
 drivers/staging/omapdrm/omap_fb.c    |   84 ++++++++++++---------
 drivers/staging/omapdrm/omap_plane.c |  136 ++++++++++++++++++++++++++++++++--
 3 files changed, 181 insertions(+), 44 deletions(-)

diff --git a/drivers/staging/omapdrm/omap_drv.h b/drivers/staging/omapdrm/omap_drv.h
index a84547c..fe4766e 100644
--- a/drivers/staging/omapdrm/omap_drv.h
+++ b/drivers/staging/omapdrm/omap_drv.h
@@ -101,8 +101,9 @@ struct drm_framebuffer *omap_framebuffer_create(struct drm_device *dev,
 struct drm_framebuffer *omap_framebuffer_init(struct drm_device *dev,
 		struct drm_mode_fb_cmd2 *mode_cmd, struct drm_gem_object **bos);
 struct drm_gem_object *omap_framebuffer_bo(struct drm_framebuffer *fb, int p);
-int omap_framebuffer_pin(struct drm_framebuffer *fb);
-void omap_framebuffer_unpin(struct drm_framebuffer *fb);
+int omap_framebuffer_replace(struct drm_framebuffer *a,
+		struct drm_framebuffer *b, void *arg,
+		void (*unpin)(void *arg, struct drm_gem_object *bo));
 void omap_framebuffer_update_scanout(struct drm_framebuffer *fb, int x, int y,
 		struct omap_overlay_info *info);
 struct drm_connector *omap_framebuffer_get_next_connector(
diff --git a/drivers/staging/omapdrm/omap_fb.c b/drivers/staging/omapdrm/omap_fb.c
index 08e2e35..fcb248f 100644
--- a/drivers/staging/omapdrm/omap_fb.c
+++ b/drivers/staging/omapdrm/omap_fb.c
@@ -137,41 +137,6 @@ static const struct drm_framebuffer_funcs omap_framebuffer_funcs = {
 	.dirty = omap_framebuffer_dirty,
 };
 
-/* pins buffer in preparation for scanout */
-int omap_framebuffer_pin(struct drm_framebuffer *fb)
-{
-	struct omap_framebuffer *omap_fb = to_omap_framebuffer(fb);
-	int ret, i, n = drm_format_num_planes(omap_fb->format->pixel_format);
-
-	for (i = 0; i < n; i++) {
-		struct plane *plane = &omap_fb->planes[i];
-		ret = omap_gem_get_paddr(plane->bo, &plane->paddr, true);
-		if (ret)
-			goto fail;
-	}
-
-	return 0;
-
-fail:
-	while (--i > 0) {
-		struct plane *plane = &omap_fb->planes[i];
-		omap_gem_put_paddr(plane->bo);
-	}
-	return ret;
-}
-
-/* releases buffer when done with scanout */
-void omap_framebuffer_unpin(struct drm_framebuffer *fb)
-{
-	struct omap_framebuffer *omap_fb = to_omap_framebuffer(fb);
-	int i, n = drm_format_num_planes(omap_fb->format->pixel_format);
-
-	for (i = 0; i < n; i++) {
-		struct plane *plane = &omap_fb->planes[i];
-		omap_gem_put_paddr(plane->bo);
-	}
-}
-
 /* update ovl info for scanout, handles cases of multi-planar fb's, etc.
  */
 void omap_framebuffer_update_scanout(struct drm_framebuffer *fb, int x, int y,
@@ -201,6 +166,55 @@ void omap_framebuffer_update_scanout(struct drm_framebuffer *fb, int x, int y,
 	}
 }
 
+/* Call for unpin 'a' (if not NULL), and pin 'b' (if not NULL).  Although
+ * buffers to unpin are just just pushed to the unpin fifo so that the
+ * caller can defer unpin until vblank.
+ *
+ * Note if this fails (ie. something went very wrong!), all buffers are
+ * unpinned, and the caller disables the overlay.  We could have tried
+ * to revert back to the previous set of pinned buffers but if things are
+ * hosed there is no guarantee that would succeed.
+ */
+int omap_framebuffer_replace(struct drm_framebuffer *a,
+		struct drm_framebuffer *b, void *arg,
+		void (*unpin)(void *arg, struct drm_gem_object *bo))
+{
+	int ret = 0, i, na, nb;
+	struct omap_framebuffer *ofba = to_omap_framebuffer(a);
+	struct omap_framebuffer *ofbb = to_omap_framebuffer(b);
+
+	na = a ? drm_format_num_planes(a->pixel_format) : 0;
+	nb = b ? drm_format_num_planes(b->pixel_format) : 0;
+
+	for (i = 0; i < max(na, nb); i++) {
+		struct plane *pa, *pb;
+
+		pa = (i < na) ? &ofba->planes[i] : NULL;
+		pb = (i < nb) ? &ofbb->planes[i] : NULL;
+
+		if (pa) {
+			unpin(arg, pa->bo);
+			pa->paddr = 0;
+		}
+
+		if (pb && !ret)
+			ret = omap_gem_get_paddr(pb->bo, &pb->paddr, true);
+	}
+
+	if (ret) {
+		/* something went wrong.. unpin what has been pinned */
+		for (i = 0; i < nb; i++) {
+			struct plane *pb = &ofba->planes[i];
+			if (pb->paddr) {
+				unpin(arg, pb->bo);
+				pb->paddr = 0;
+			}
+		}
+	}
+
+	return ret;
+}
+
 struct drm_gem_object *omap_framebuffer_bo(struct drm_framebuffer *fb, int p)
 {
 	struct omap_framebuffer *omap_fb = to_omap_framebuffer(fb);
diff --git a/drivers/staging/omapdrm/omap_plane.c b/drivers/staging/omapdrm/omap_plane.c
index c5625e3..55ddc58 100644
--- a/drivers/staging/omapdrm/omap_plane.c
+++ b/drivers/staging/omapdrm/omap_plane.c
@@ -17,6 +17,8 @@
  * this program.  If not, see <http://www.gnu.org/licenses/>.
  */
 
+#include <linux/kfifo.h>
+
 #include "omap_drv.h"
 
 /* some hackery because omapdss has an 'enum omap_plane' (which would be
@@ -46,8 +48,57 @@ struct omap_plane {
 
 	uint32_t nformats;
 	uint32_t formats[32];
+
+	/* for synchronizing access to unpins fifo */
+	struct mutex unpin_mutex;
+
+	/* set of bo's pending unpin until next END_WIN irq */
+	DECLARE_KFIFO_PTR(unpin_fifo, struct drm_gem_object *);
+	int num_unpins, pending_num_unpins;
+
+	/* for deferred unpin when we need to wait for scanout complete irq */
+	struct work_struct work;
+};
+
+/* map from ovl->id to the irq we are interested in for scanout-done */
+static const uint32_t id2irq[] = {
+		[OMAP_DSS_GFX]    = DISPC_IRQ_GFX_END_WIN,
+		[OMAP_DSS_VIDEO1] = DISPC_IRQ_VID1_END_WIN,
+		[OMAP_DSS_VIDEO2] = DISPC_IRQ_VID2_END_WIN,
+		[OMAP_DSS_VIDEO3] = DISPC_IRQ_VID3_END_WIN,
 };
 
+static void dispc_isr(void *arg, uint32_t mask)
+{
+	struct drm_plane *plane = arg;
+	struct omap_plane *omap_plane = to_omap_plane(plane);
+	struct omap_drm_private *priv = plane->dev->dev_private;
+
+	omap_dispc_unregister_isr(dispc_isr, plane,
+			id2irq[omap_plane->ovl->id]);
+
+	queue_work(priv->wq, &omap_plane->work);
+}
+
+static void unpin_worker(struct work_struct *work)
+{
+	struct omap_plane *omap_plane =
+			container_of(work, struct omap_plane, work);
+
+	mutex_lock(&omap_plane->unpin_mutex);
+	DBG("unpinning %d of %d", omap_plane->num_unpins,
+			omap_plane->num_unpins + omap_plane->pending_num_unpins);
+	while (omap_plane->num_unpins > 0) {
+		struct drm_gem_object *bo = NULL;
+		int ret = kfifo_get(&omap_plane->unpin_fifo, &bo);
+		WARN_ON(!ret);
+		omap_gem_put_paddr(bo);
+		drm_gem_object_unreference_unlocked(bo);
+		omap_plane->num_unpins--;
+	}
+	mutex_unlock(&omap_plane->unpin_mutex);
+}
+
 /* push changes down to dss2 */
 static int commit(struct drm_plane *plane)
 {
@@ -73,6 +124,11 @@ static int commit(struct drm_plane *plane)
 		return ret;
 	}
 
+	mutex_lock(&omap_plane->unpin_mutex);
+	omap_plane->num_unpins += omap_plane->pending_num_unpins;
+	omap_plane->pending_num_unpins = 0;
+	mutex_unlock(&omap_plane->unpin_mutex);
+
 	/* our encoder doesn't necessarily get a commit() after this, in
 	 * particular in the dpms() and mode_set_base() cases, so force the
 	 * manager to update:
@@ -85,8 +141,29 @@ static int commit(struct drm_plane *plane)
 			dev_err(dev->dev, "could not apply settings\n");
 			return ret;
 		}
+
+		/*
+		 * NOTE: really this should be atomic w/ mgr->apply() but
+		 * omapdss does not expose such an API
+		 */
+		if (omap_plane->num_unpins > 0) {
+			ret = omap_dispc_register_isr(dispc_isr,
+				plane, id2irq[ovl->id]);
+		}
+
+		/*
+		 * omapdss has upper limit on # of registered irq handlers,
+		 * which we shouldn't hit.. but if we do the limit should
+		 * be raised or bad things happen:
+		 */
+		WARN_ON(ret == -EBUSY);
+
+	} else {
+		struct omap_drm_private *priv = dev->dev_private;
+		queue_work(priv->wq, &omap_plane->work);
 	}
 
+
 	if (ovl->is_enabled(ovl)) {
 		omap_framebuffer_flush(plane->fb, info->pos_x, info->pos_y,
 				info->out_width, info->out_height);
@@ -139,21 +216,48 @@ static void update_manager(struct drm_plane *plane)
 	}
 }
 
+static void unpin(void *arg, struct drm_gem_object *bo)
+{
+	struct drm_plane *plane = arg;
+	struct omap_plane *omap_plane = to_omap_plane(plane);
+
+	if (kfifo_put(&omap_plane->unpin_fifo,
+			(const struct drm_gem_object **)&bo)) {
+		omap_plane->pending_num_unpins++;
+		/* also hold a ref so it isn't free'd while pinned */
+		drm_gem_object_reference(bo);
+	} else {
+		dev_err(plane->dev->dev, "unpin fifo full!\n");
+		omap_gem_put_paddr(bo);
+	}
+}
+
 /* update which fb (if any) is pinned for scanout */
 static int update_pin(struct drm_plane *plane, struct drm_framebuffer *fb)
 {
 	struct omap_plane *omap_plane = to_omap_plane(plane);
-	int ret = 0;
+	struct drm_framebuffer *pinned_fb = omap_plane->pinned_fb;
+
+	if (pinned_fb != fb) {
+		int ret;
+
+		DBG("%p -> %p", pinned_fb, fb);
+
+		mutex_lock(&omap_plane->unpin_mutex);
+		ret = omap_framebuffer_replace(pinned_fb, fb, plane, unpin);
+		mutex_unlock(&omap_plane->unpin_mutex);
+
+		if (ret) {
+			dev_err(plane->dev->dev, "could not swap %p -> %p\n",
+					omap_plane->pinned_fb, fb);
+			omap_plane->pinned_fb = NULL;
+			return ret;
+		}
 
-	if (omap_plane->pinned_fb != fb) {
-		if (omap_plane->pinned_fb)
-			omap_framebuffer_unpin(omap_plane->pinned_fb);
 		omap_plane->pinned_fb = fb;
-		if (fb)
-			ret = omap_framebuffer_pin(fb);
 	}
 
-	return ret;
+	return 0;
 }
 
 /* update parameters that are dependent on the framebuffer dimensions and
@@ -243,6 +347,8 @@ static void omap_plane_destroy(struct drm_plane *plane)
 	DBG("%s", omap_plane->ovl->name);
 	omap_plane_disable(plane);
 	drm_plane_cleanup(plane);
+	WARN_ON(omap_plane->pending_num_unpins + omap_plane->num_unpins > 0);
+	kfifo_free(&omap_plane->unpin_fifo);
 	kfree(omap_plane);
 }
 
@@ -260,8 +366,10 @@ int omap_plane_dpms(struct drm_plane *plane, int mode)
 		if (!r)
 			r = ovl->enable(ovl);
 	} else {
+		struct omap_drm_private *priv = plane->dev->dev_private;
 		r = ovl->disable(ovl);
 		update_pin(plane, NULL);
+		queue_work(priv->wq, &omap_plane->work);
 	}
 
 	return r;
@@ -280,16 +388,30 @@ struct drm_plane *omap_plane_init(struct drm_device *dev,
 {
 	struct drm_plane *plane = NULL;
 	struct omap_plane *omap_plane;
+	int ret;
 
 	DBG("%s: possible_crtcs=%08x, priv=%d", ovl->name,
 			possible_crtcs, priv);
 
+	/* friendly reminder to update table for future hw: */
+	WARN_ON(ovl->id >= ARRAY_SIZE(id2irq));
+
 	omap_plane = kzalloc(sizeof(*omap_plane), GFP_KERNEL);
 	if (!omap_plane) {
 		dev_err(dev->dev, "could not allocate plane\n");
 		goto fail;
 	}
 
+	mutex_init(&omap_plane->unpin_mutex);
+
+	ret = kfifo_alloc(&omap_plane->unpin_fifo, 16, GFP_KERNEL);
+	if (ret) {
+		dev_err(dev->dev, "could not allocate unpin FIFO\n");
+		goto fail;
+	}
+
+	INIT_WORK(&omap_plane->work, unpin_worker);
+
 	omap_plane->nformats = omap_framebuffer_get_formats(
 			omap_plane->formats, ARRAY_SIZE(omap_plane->formats),
 			ovl->supported_modes);
-- 
1.7.5.4