drm/vc4: Use interrupts for frame end handling.

We still exec synchronously, but this gets rid of the busy wait.

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

3 files changed, 44 insertions, 12 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index 8b745c3bac40..4c600e0d3a85 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -19,6 +19,9 @@ struct vc4_dev {
 	struct vc4_crtc *crtc[3];
 	struct vc4_v3d *v3d;
 
+	wait_queue_head_t frame_done_queue;
+	bool frame_done;
+
 	/* List of struct vc4_list_bo_entry allocated to accomodate
 	 * binner overflow.  These will be freed when the exec is
 	 * done.
diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index 3c38e41bd41f..35e14f81ddbf 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -61,21 +61,37 @@ submit_cl(struct drm_device *dev, uint32_t thread, uint32_t start, uint32_t end)
 	barrier();
 }
 
-static bool
-thread_stopped(struct drm_device *dev, uint32_t thread)
+static int
+vc4_wait_for_job(struct drm_device *dev, struct exec_info *exec,
+		 unsigned timeout)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int ret = 0;
+	unsigned long timeout_expire;
+	DEFINE_WAIT(wait);
 
-	barrier();
-	return !(V3D_READ(V3D_PCS) & (0x3 << thread));
-}
+	if (vc4->frame_done)
+		return 0;
 
-static int
-wait_for_job(struct drm_device *dev, struct exec_info *exec)
-{
-	int ret;
+	timeout_expire = jiffies + msecs_to_jiffies(timeout);
+
+	for (;;) {
+		prepare_to_wait(&vc4->frame_done_queue, &wait,
+				TASK_UNINTERRUPTIBLE);
+
+		if (time_after_eq(jiffies, timeout_expire)) {
+			ret = -ETIME;
+			break;
+		}
+
+		if (vc4->frame_done)
+			break;
+
+		schedule_timeout(timeout_expire - jiffies);
+	}
+
+	finish_wait(&vc4->frame_done_queue, &wait);
 
-	ret = wait_for(thread_stopped(dev, 1), 1000);
 	if (ret) {
 		DRM_ERROR("timeout waiting for render thread idle\n");
 		return ret;
@@ -125,10 +141,11 @@ vc4_submit(struct drm_device *dev, struct exec_info *exec)
 	V3D_WRITE(V3D_BPOA, 0);
 	V3D_WRITE(V3D_BPOS, 0);
 
+	vc4->frame_done = false;
 	submit_cl(dev, 0, ct0ca, ct0ea);
 	submit_cl(dev, 1, ct1ca, ct1ea);
 
-	ret = wait_for_job(dev, exec);
+	ret = vc4_wait_for_job(dev, exec, 1000);
 	if (ret)
 		return ret;
 
diff --git a/drivers/gpu/drm/vc4/vc4_irq.c b/drivers/gpu/drm/vc4/vc4_irq.c
index dfc302606762..b0e6fa847367 100644
--- a/drivers/gpu/drm/vc4/vc4_irq.c
+++ b/drivers/gpu/drm/vc4/vc4_irq.c
@@ -24,7 +24,10 @@
 #include "vc4_drv.h"
 #include "vc4_regs.h"
 
-#define V3D_DRIVER_IRQS V3D_INT_OUTOMEM
+#define V3D_DRIVER_IRQS (V3D_INT_OUTOMEM | \
+			 V3D_INT_FRDONE)
+
+DECLARE_WAIT_QUEUE_HEAD(render_wait);
 
 static void
 vc4_overflow_mem_work(struct work_struct *work)
@@ -70,6 +73,11 @@ vc4_irq(int irq, void *arg)
 		schedule_work(&vc4->overflow_mem_work);
 	}
 
+	if (intctl & V3D_INT_FRDONE) {
+		vc4->frame_done = true;
+		wake_up_all(&vc4->frame_done_queue);
+	}
+
 	return intctl ? IRQ_HANDLED : IRQ_NONE;
 }
 
@@ -78,6 +86,7 @@ vc4_irq_preinstall(struct drm_device *dev)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 
+	init_waitqueue_head(&vc4->frame_done_queue);
 	INIT_WORK(&vc4->overflow_mem_work, vc4_overflow_mem_work);
 
 	/* Clear any pending interrupts someone might have left around
@@ -126,4 +135,7 @@ void vc4_irq_reset(struct drm_device *dev)
 	V3D_WRITE(V3D_INTCTL, V3D_DRIVER_IRQS);
 	V3D_WRITE(V3D_INTDIS, 0);
 	V3D_WRITE(V3D_INTENA, V3D_DRIVER_IRQS);
+
+	vc4->frame_done = true;
+	wake_up_all(&vc4->frame_done_queue);
 }