[PATCH v2 09/11] video: lcdif: drain write-combine framebuffer in fb_damage

[Johannes Schneider]

From: Thomas Haemmerle <thomas.haemmerle at leica-geosystems.com>

The framebuffer is allocated Normal Non-Cacheable (write-combine).
After the splash command's memcpy() from a cached shadow buffer into
the hardware framebuffer (see gu_screen_blit()), stores may still
reside in the CPU write-combine buffer when scanout next fetches --
visible as partial / corrupted rendering at the bottom of the image.

dsb() alone is not sufficient: it orders subsequent instructions
against the outstanding writes but doesn't actively drain the WC
buffer to memory.  A readback of any byte inside the WC region does,
since the load cannot complete until prior stores to the buffer have
been retired.

Wire lcdif_fb_damage to do that readback after each blit.  Keep
fb_flush bound to lcdif_crtc_atomic_flush so the SHADOW_LOAD_EN
re-latch still fires on fb_close / fbtest paths.

Suggested-by: Lucas Stach <l.stach at pengutronix.de>
Signed-off-by: Thomas Haemmerle <thomas.haemmerle at leica-geosystems.com>
---
 drivers/video/lcdif_kms.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/drivers/video/lcdif_kms.c b/drivers/video/lcdif_kms.c
index e3d1985dc9..57ca92b4be 100644
--- a/drivers/video/lcdif_kms.c
+++ b/drivers/video/lcdif_kms.c
@@ -396,10 +396,17 @@ static void lcdif_disable_fb_controller(struct fb_info *info)
 	lcdif_disable_controller(lcdif);
 }
 
+static void lcdif_fb_damage(struct fb_info *info, const struct fb_rect *rect)
+{
+	/* readback drains the write-combine buffer (dsb() doesn't) */
+	(void)*(volatile u8 *)info->screen_base;
+}
+
 static struct fb_ops lcdif_fb_ops = {
 	.fb_enable = lcdif_enable_fb_controller,
 	.fb_disable = lcdif_disable_fb_controller,
 	.fb_flush = lcdif_crtc_atomic_flush,
+	.fb_damage = lcdif_fb_damage,
 };
 
 /* -----------------------------------------------------------------------------
-- 
2.43.0