Re: [PATCH v6 07/15] drm/rockchip: vop2: Reset AXI and DCLK to improve robustness

From: Tim Surber

Date: Fri Nov 20 2026 - 21:32:26 EST


Hi Cristian,

On Thu, 05 Nov 2026 03:00:41 +0200, Cristian Ciocaltea wrote:
> @@ -953,6 +974,8 @@ static int vop2_enable(struct vop2 *vop2)
>
> static void vop2_disable(struct vop2 *vop2)
> {
> + vop2_reset_assert_deassert(vop2, vop2->axi_rst);
> +
> rockchip_drm_dma_detach_device(vop2->drm, vop2->dev);

with this patch my RK3588 board (FriendlyElec
CM3588) hangs hard sometimes when the last CRTC is switched off. Without the series
or with the AXI reset in vop2_disable() removed it does not happen.

Here is a minimal reproducer, run as root, with a display on HDMI and fbcon on it:

python3 - <<'EOF' &
import fcntl, os, struct
fd = os.open("/dev/dri/by-path/platform-display-subsystem-card", os.O_RDWR)
try:
fcntl.ioctl(fd, 0x641F) # DRM_IOCTL_DROP_MASTER: leave KMS to fbcon
except OSError:
pass
while True:
b = bytearray(struct.pack("6IQ", 2160, 3840, 32, 0, 0, 0, 0))
fcntl.ioctl(fd, 0xC02064B2, b) # DRM_IOCTL_MODE_CREATE_DUMB
handle = struct.unpack("6IQ", b)[4]
fcntl.ioctl(fd, 0xC00464B4, struct.pack("I", handle)) # DRM_IOCTL_MODE_DESTROY_DUMB
EOF
while :; do
echo 4 > /sys/class/graphics/fb0/blank; sleep 0.3
echo 0 > /sys/class/graphics/fb0/blank; sleep 0.7
done

This hangs the board within a few blank cycles. In
practice I hit it when a KMS application exits: it frees its buffers
while fbcon takes the display back with a full modeset.

This issue can be fixed for me by moving vop2_reset_assert_deassert(vop2, vop2->axi_rst) from vop2_disable to vop2_enable.

--8<--
diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c b/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
--- a/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
+++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
@@ -913,6 +913,9 @@ static int vop2_enable(struct vop2 *vop2)
if (ret)
goto err_put_pm;

+ /* Reset while idle; doing it in vop2_disable() can hang the SoC */
+ vop2_reset_assert_deassert(vop2, vop2->axi_rst);
+
ret = rockchip_drm_dma_attach_device(vop2->drm, vop2->dev);
if (ret) {
drm_err(vop2->drm, "failed to attach dma mapping: %d\n", ret);
@@ -973,8 +976,6 @@ static int vop2_enable(struct vop2 *vop2)

static void vop2_disable(struct vop2 *vop2)
{
- vop2_reset_assert_deassert(vop2, vop2->axi_rst);
-
rockchip_drm_dma_detach_device(vop2->drm, vop2->dev);

regcache_drop_region(vop2->map, 0, vop2_regmap_config.max_register);
-->8--

Does that work for the cases the reset was added for (the RGB <-> YUV420
switch glitches and the DSI panel on the EVB1)? I can't test those here.


Thanks,
Tim