On Thu, Apr 21, 2016 at 09:41:34AM +0200, Daniel Vetter wrote:

On Wed, Apr 20, 2016 at 09:04:38PM +0200, Noralf Trønnes wrote:

quoted

Den 20.04.2016 19:47, skrev Daniel Vetter:

quoted

On Wed, Apr 20, 2016 at 05:25:28PM +0200, Noralf Trønnes wrote:

quoted

Use the fbdev deferred io support in drm_fb_helper.
The (struct fb_ops *)->fb_{fillrect,copyarea,imageblit} functions will
now be deferred in the same way that mmap damage is, instead of being
flushed directly.
This patch has only been compile tested.

Signed-off-by: Noralf Trønnes <redacted>
---
 drivers/gpu/drm/qxl/qxl_display.c |   9 +-
 drivers/gpu/drm/qxl/qxl_drv.h     |   7 +-
 drivers/gpu/drm/qxl/qxl_fb.c      | 220 ++++++++++----------------------------
 drivers/gpu/drm/qxl/qxl_kms.c     |   4 -
 4 files changed, 62 insertions(+), 178 deletions(-)

diff --git a/drivers/gpu/drm/qxl/qxl_display.c b/drivers/gpu/drm/qxl/qxl_display.c
index 030409a..9a03524 100644
--- a/drivers/gpu/drm/qxl/qxl_display.c
+++ b/drivers/gpu/drm/qxl/qxl_display.c

@@ -465,7 +465,7 @@ static const struct drm_crtc_funcs qxl_crtc_funcs = {
 	.page_flip = qxl_crtc_page_flip,
 };
-static void qxl_user_framebuffer_destroy(struct drm_framebuffer *fb)
+void qxl_user_framebuffer_destroy(struct drm_framebuffer *fb)
 {
 	struct qxl_framebuffer *qxl_fb = to_qxl_framebuffer(fb);

@@ -527,12 +527,13 @@ int
 qxl_framebuffer_init(struct drm_device *dev,
 		     struct qxl_framebuffer *qfb,
 		     const struct drm_mode_fb_cmd2 *mode_cmd,
-		     struct drm_gem_object *obj)
+		     struct drm_gem_object *obj,
+		     const struct drm_framebuffer_funcs *funcs)

There should be no need at all to have a separate fb funcs table for the
fbdev fb. Both /should/ be able to use the exact same (already existing)
->dirty() callback. We need this only in CMA because CMA is a midlayer
used by multiple drivers.

I don't see how I can avoid it.

fbdev framebuffer flushing:

static void qxl_fb_dirty_flush(struct fb_info *info)
{
        qxl_fb_image_init(&qxl_fb_image, qdev, info, NULL);
        qxl_draw_opaque_fb(&qxl_fb_image, stride);
}

drm framebuffer flushing:

static int qxl_framebuffer_surface_dirty(...)
{
        qxl_draw_dirty_fb(...);
}

qxl_draw_opaque_fb() and qxl_draw_dirty_fb() differ so much that it's way
over my head to see if they can be combined.
Here's an online diff of the two functions:
https://www.diffchecker.com/jqbbalux

Imo nuke the fbdev one entirely. If it breaks then it's either a bug in
your generic fbdefio code, or the qxl ->dirty implementation has a bug. It
should work ;-)

Ok, slightly more seriously the difference seems to be that the fbdev one
support paletted mode too. But since qxl has 0 pixel format checking
anywhere I have no idea whether that's dead code (i.e. broken) or actually
working. I guess keeping the split is ok, if we add a big FIXME comment to
it that this is very fishy.

Ok, I read around a bit more. The only things qxl seems to support are
bits_per_pixel of 1, 24 and 32 (see qxl_image_init_helper). And drm has no
way to pass in 1 bpp images. And it doesn't support 8 bit paletted, which
is the only paletted thing drm supports.

So if you totally feel like I think we could add format checking for
DRM_FORMAT_XRGB8888 and DRM_FORMAT_RGB888 in qxl_framebuffer_init and then
rip out all that code. But that's a few more patches and probably should
be tested actually ;-)

FIXME plus explaing it all in the commit message is fine with me too.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch