[PATCH 3/3] drm/vc4: crtc: Reduce PV fifo threshold on hvs4

Maxime Ripard maxime at cerno.tech
Thu Mar 18 16:13:28 UTC 2021


From: Dom Cobley <popcornmix at gmail.com>

Experimentally have found PV on hvs4 reports fifo full
error with expected settings and does not with one less

This appears as:
[drm:drm_atomic_helper_wait_for_flip_done] *ERROR* [CRTC:82:crtc-3] flip_done timed out

with bit 10 of PV_STAT set "HVS driving pixels when the PV FIFO is full"

Fixes: c8b75bca92cb ("drm/vc4: Add KMS support for Raspberry Pi.")
Signed-off-by: Dom Cobley <popcornmix at gmail.com>
Signed-off-by: Maxime Ripard <maxime at cerno.tech>
---
 drivers/gpu/drm/vc4/vc4_crtc.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

diff --git a/drivers/gpu/drm/vc4/vc4_crtc.c b/drivers/gpu/drm/vc4/vc4_crtc.c
index 269390bc586e..76657dcdf9b0 100644
--- a/drivers/gpu/drm/vc4/vc4_crtc.c
+++ b/drivers/gpu/drm/vc4/vc4_crtc.c
@@ -210,6 +210,7 @@ static u32 vc4_get_fifo_full_level(struct vc4_crtc *vc4_crtc, u32 format)
 {
 	const struct vc4_crtc_data *crtc_data = vc4_crtc_to_vc4_crtc_data(vc4_crtc);
 	const struct vc4_pv_data *pv_data = vc4_crtc_to_vc4_pv_data(vc4_crtc);
+	struct vc4_dev *vc4 = to_vc4_dev(vc4_crtc->base.dev);
 	u32 fifo_len_bytes = pv_data->fifo_depth;
 
 	/*
@@ -238,6 +239,22 @@ static u32 vc4_get_fifo_full_level(struct vc4_crtc *vc4_crtc, u32 format)
 		if (crtc_data->hvs_output == 5)
 			return 32;
 
+		/*
+		 * It looks like in some situations, we will overflow
+		 * the PixelValve FIFO (with the bit 10 of PV stat being
+		 * set) and stall the HVS / PV, eventually resulting in
+		 * a page flip timeout.
+		 *
+		 * Displaying the video overlay during a playback with
+		 * Kodi on an RPi3 seems to be a great solution with a
+		 * failure rate around 50%.
+		 *
+		 * Removing 1 from the FIFO full level however
+		 * seems to completely remove that issue.
+		 */
+		if (!vc4->hvs->hvs5)
+			return fifo_len_bytes - 3 * HVS_FIFO_LATENCY_PIX - 1;
+
 		return fifo_len_bytes - 3 * HVS_FIFO_LATENCY_PIX;
 	}
 }
-- 
2.30.2



More information about the dri-devel mailing list