drm/vc4: Add support for async pageflips.

This should be the fastest screen update path we have, but it doesn't
seem to have really improved performance with vblank_mode=0
fullscreen.

Signed-off-by: Eric Anholt <eric@anholt.net>

5 files changed, 140 insertions, 7 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_crtc.c b/drivers/gpu/drm/vc4/vc4_crtc.c
index f9960ac34fd7..446f8b3aa436 100644
--- a/drivers/gpu/drm/vc4/vc4_crtc.c
+++ b/drivers/gpu/drm/vc4/vc4_crtc.c
@@ -10,8 +10,10 @@
  * Controls the timings of the hardware's pixel valve.
  */
 
+#include "drm_atomic.h"
 #include "drm_atomic_helper.h"
 #include "drm_crtc_helper.h"
+#include "drm_fb_cma_helper.h"
 #include "linux/component.h"
 #include "vc4_drv.h"
 #include "vc4_regs.h"
@@ -329,10 +331,105 @@ static irqreturn_t vc4_crtc_irq_handler(int irq, void *data)
 	return ret;
 }
 
+struct vc4_async_flip_state {
+	struct drm_crtc *crtc;
+	struct drm_framebuffer *fb;
+	struct drm_pending_vblank_event *event;
+
+	struct vc4_seqno_cb cb;
+};
+
+/* Called when the V3D execution for the BO being flipped to is done, so that
+ * we can actually update the plane's address to point to it.
+ */
+static void
+vc4_async_page_flip_complete(struct vc4_seqno_cb *cb)
+{
+	struct vc4_async_flip_state *flip_state =
+		container_of(cb, struct vc4_async_flip_state, cb);
+	struct drm_crtc *crtc = flip_state->crtc;
+	struct drm_device *dev = crtc->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct drm_plane *plane = crtc->primary;
+
+	vc4_plane_async_set_fb(plane, flip_state->fb);
+	if (flip_state->event) {
+		unsigned long flags;
+		spin_lock_irqsave(&dev->event_lock, flags);
+		drm_crtc_send_vblank_event(crtc, flip_state->event);
+		spin_unlock_irqrestore(&dev->event_lock, flags);
+	}
+
+	drm_framebuffer_unreference(flip_state->fb);
+	kfree(flip_state);
+
+	up(&vc4->async_modeset);
+}
+
+/* Implements async (non-vblank-synced) page flips.
+ *
+ * The page flip ioctl needs to return immediately, so we grab the
+ * modeset semaphore on the pipe, and queue the address update for
+ * when V3D is done with the BO being flipped to.
+ */
+static int vc4_async_page_flip(struct drm_crtc *crtc,
+			       struct drm_framebuffer *fb,
+			       struct drm_pending_vblank_event *event,
+			       uint32_t flags)
+{
+	struct drm_device *dev = crtc->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct drm_plane *plane = crtc->primary;
+	int ret = 0;
+	struct vc4_async_flip_state *flip_state;
+	struct drm_gem_cma_object *cma_bo = drm_fb_cma_get_gem_obj(fb, 0);
+	struct vc4_bo *bo = to_vc4_bo(&cma_bo->base);
+
+	flip_state = kzalloc(sizeof(*flip_state), GFP_KERNEL);
+	if (!flip_state)
+		return -ENOMEM;
+
+	drm_framebuffer_reference(fb);
+	flip_state->fb = fb;
+	flip_state->crtc = crtc;
+	flip_state->event = event;
+
+	/* Make sure all other async modesetes have landed. */
+	ret = down_interruptible(&vc4->async_modeset);
+	if (ret) {
+		kfree(flip_state);
+		return ret;
+	}
+
+	/* Immediately update the plane's legacy fb pointer, so that later
+	 * modeset prep sees the state that will be present when the semaphore
+	 * is released.
+	 */
+	drm_atomic_set_fb_for_plane(plane->state, fb);
+	plane->fb = fb;
+
+	vc4_queue_seqno_cb(dev, &flip_state->cb, bo->seqno,
+			   vc4_async_page_flip_complete);
+
+	/* Driver takes ownership of state on successful async commit. */
+	return 0;
+}
+
+static int vc4_page_flip(struct drm_crtc *crtc,
+		  struct drm_framebuffer *fb,
+		  struct drm_pending_vblank_event *event,
+		  uint32_t flags)
+{
+	if (flags & DRM_MODE_PAGE_FLIP_ASYNC)
+		return vc4_async_page_flip(crtc, fb, event, flags);
+	else
+		return drm_atomic_helper_page_flip(crtc, fb, event, flags);
+}
+
 static const struct drm_crtc_funcs vc4_crtc_funcs = {
 	.set_config = drm_atomic_helper_set_config,
 	.destroy = vc4_crtc_destroy,
-	.page_flip = drm_atomic_helper_page_flip,
+	.page_flip = vc4_page_flip,
 	.set_property = NULL,
 	.cursor_set = NULL, /* handled by drm_mode_cursor_universal */
 	.cursor_move = NULL, /* handled by drm_mode_cursor_universal */
diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index 13b475785ca0..c2db6bd57352 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -431,6 +431,7 @@ struct drm_plane *vc4_plane_init(struct drm_device *dev,
 				 enum drm_plane_type type);
 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
 u32 vc4_plane_dlist_size(struct drm_plane_state *state);
+void vc4_plane_async_set_fb(struct drm_plane *plane, struct drm_framebuffer *fb);
 
 /* vc4_v3d.c */
 void vc4_v3d_register(void);
diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index 577cc6d3840f..ca3955c5ea79 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -636,12 +636,6 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 
 	mutex_unlock(&dev->struct_mutex);
 
-	/* To keep any client from getting too far ahead (particularly
-	 * a problem when BO caching is involved), we wait on the
-	 * previous rendering before returning to userspace.
-	 */
-	vc4_wait_for_seqno(dev, args->seqno - 1, ~0ull, true);
-
 	return 0;
 
 fail:
diff --git a/drivers/gpu/drm/vc4/vc4_kms.c b/drivers/gpu/drm/vc4/vc4_kms.c
index acfd1b23ab4a..44836d51145d 100644
--- a/drivers/gpu/drm/vc4/vc4_kms.c
+++ b/drivers/gpu/drm/vc4/vc4_kms.c
@@ -207,6 +207,7 @@ vc4_kms_load(struct drm_device *dev)
 	dev->mode_config.max_height = 2048;
 	dev->mode_config.funcs = &vc4_mode_funcs;
 	dev->mode_config.preferred_depth = 24;
+	dev->mode_config.async_page_flip = true;
 
 	ret = vc4_init_modeset_objects(dev);
 	if (ret)
diff --git a/drivers/gpu/drm/vc4/vc4_plane.c b/drivers/gpu/drm/vc4/vc4_plane.c
index 0c8afb17314a..0d703154b29d 100644
--- a/drivers/gpu/drm/vc4/vc4_plane.c
+++ b/drivers/gpu/drm/vc4/vc4_plane.c
@@ -21,6 +21,14 @@ struct vc4_plane_state {
 	u32 *dlist;
 	u32 dlist_size; /* Number of dwords in allocated for the display list */
 	u32 dlist_count; /* Number of used dwords in the display list. */
+
+	/* Offset in the dlist to pointer word 0. */
+	u32 pw0_offset;
+
+	/* Offset where the plane's dlist was last stored in the
+	   hardware at vc4_crtc_atomic_flush() time.
+	*/
+	u32 *hw_dlist;
 };
 
 static inline struct vc4_plane_state *
@@ -191,6 +199,8 @@ vc4_plane_mode_set(struct drm_plane *plane, struct drm_plane_state *state)
 	/* Position Word 3: Context.  Written by the HVS. */
 	vc4_dlist_write(vc4_state, 0xc0c0c0c0);
 
+	vc4_state->pw0_offset = vc4_state->dlist_count;
+
 	/* Pointer Word 0: RGB / Y Pointer */
 	vc4_dlist_write(vc4_state, bo->paddr + offset);
 
@@ -244,6 +254,8 @@ u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist)
 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
 	int i;
 
+	vc4_state->hw_dlist = dlist;
+
 	/* Can't memcpy_toio() because it needs to be 32-bit writes. */
 	for (i = 0; i < vc4_state->dlist_count; i++)
 		writel(vc4_state->dlist[i], &dlist[i]);
@@ -258,6 +270,34 @@ u32 vc4_plane_dlist_size(struct drm_plane_state *state)
 	return vc4_state->dlist_count;
 }
 
+/* Updates the plane to immediately (well, once the FIFO needs
+ * refilling) scan out from at a new framebuffer.
+ */
+void vc4_plane_async_set_fb(struct drm_plane *plane, struct drm_framebuffer *fb)
+{
+	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
+	struct drm_gem_cma_object *bo = drm_fb_cma_get_gem_obj(fb, 0);
+	uint32_t addr;
+
+	/* We're skipping the address adjustment for negative origin,
+	 * because this is only called on the primary plane.
+	 */
+	WARN_ON_ONCE(plane->state->crtc_x < 0 || plane->state->crtc_y < 0);
+	addr = bo->paddr + fb->offsets[0];
+
+	/* Write the new address into the hardware immediately.  The
+	 * scanout will start from this address as soon as the FIFO
+	 * needs to refill with pixels.
+	 */
+	writel(addr, &vc4_state->hw_dlist[vc4_state->pw0_offset]);
+
+	/* Also update the CPU-side dlist copy, so that any later
+	 * atomic updates that don't do a new modeset on our plane
+	 * also use our updated address.
+	 */
+	vc4_state->dlist[vc4_state->pw0_offset] = addr;
+}
+
 static const struct drm_plane_helper_funcs vc4_plane_helper_funcs = {
 	.prepare_fb = NULL,
 	.cleanup_fb = NULL,