Re: [PATCH v2 4/8] drm/fb-helper: Add fb_deferred_io support

On Wed, Apr 27, 2016 at 11:45:31AM +0200, Noralf Trønnes wrote:

Den 26.04.2016 19:19, skrev Daniel Vetter:

quoted

On Tue, Apr 26, 2016 at 06:24:54PM +0200, Noralf Trønnes wrote:

quoted

Den 25.04.2016 11:09, skrev Daniel Vetter:

quoted

On Sun, Apr 24, 2016 at 10:48:58PM +0200, Noralf Trønnes wrote:

quoted

This adds deferred io support if CONFIG_FB_DEFERRED_IO is enabled.
The fbdev framebuffer changes are flushed using the callback
(struct drm_framebuffer *)->funcs->dirty() by a dedicated worker
ensuring that it always runs in process context.

Signed-off-by: Noralf Trønnes <redacted>
---

Changes since v1:
- Use a dedicated worker to run the framebuffer flushing like qxl does
- Add parameter descriptions to drm_fb_helper_deferred_io

 drivers/gpu/drm/drm_fb_helper.c | 127 +++++++++++++++++++++++++++++++++++++++-
 include/drm/drm_fb_helper.h     |  17 ++++++
 2 files changed, 143 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
index 855108e..46ee6f8 100644
--- a/drivers/gpu/drm/drm_fb_helper.c
+++ b/drivers/gpu/drm/drm_fb_helper.c

@@ -40,6 +40,7 @@
 #include <drm/drm_crtc_helper.h>
 #include <drm/drm_atomic.h>
 #include <drm/drm_atomic_helper.h>
+#include <drm/drm_rect.h>

 static bool drm_fbdev_emulation = true;
 module_param_named(fbdev_emulation, drm_fbdev_emulation, bool, 0600);

@@ -48,6 +49,10 @@ MODULE_PARM_DESC(fbdev_emulation,

 static LIST_HEAD(kernel_fb_helper_list);

+static void drm_fb_helper_dirty_init(struct drm_fb_helper *helper);
+static void drm_fb_helper_dirty(struct fb_info *info, u32 x, u32 y,
+				u32 width, u32 height);
+
 /**
  * DOC: fbdev helpers
  *

@@ -84,6 +89,16 @@ static LIST_HEAD(kernel_fb_helper_list);
  * and set up an initial configuration using the detected hardware, drivers
  * should call drm_fb_helper_single_add_all_connectors() followed by
  * drm_fb_helper_initial_config().
+ *
+ * If CONFIG_FB_DEFERRED_IO is enabled and
+ * (struct drm_framebuffer *)->funcs->dirty is set, the
+ * drm_fb_helper_{cfb,sys}_{write,fillrect,copyarea,imageblit} functions
+ * will accumulate changes and schedule (struct fb_helper).dirty_work to run
+ * right away. This worker then calls the dirty() function ensuring that it
+ * will always run in process context since the fb_*() function could be
+ * running in atomic context. If drm_fb_helper_deferred_io() is used as the
+ * deferred_io callback it will also schedule dirty_work with the damage
+ * collected from the mmap page writes.

One thing to consider (and personally I don't care either way) is whether
we shouldn't just select CONFIG_FB_DEFERRED_IO if the fbdev helpers are
enabled. Pushing that out to drivers is imo a bit fragile.

But like I said I'm ok with either way.

My concern was adding code and data that only a few drivers would
actually use. But of course there's the tradeoff with complexity.
I use this to enable it:
        select FB_DEFERRED_IO if DRM_KMS_FB_HELPER

I guess the maintainer has to make this choice between size and complexity
:-)
I can enable it by default if you want, drm is both huge and complex so I
don't know what's best.

As a sidenote, I have also put all the fbdev code in a file of it's own to
make it simple with regards to the DRM_FBDEV_EMULATION user option:
tinydrm-$(CONFIG_DRM_KMS_FB_HELPER)     += tinydrm-fbdev.o

Ok, if you ask maintainers then please nuke the #ifdef from .c files. If
you select CONFIG_DRM_KMS_FB_HELPER, then you get hdmi, edid, dp aux, dp
mst and whatever else helpers, even if you don't need them. Adding 3
functions for defio when you select fbdev helpers and maybe don't need
them is totally harmless. And removing the #ifdef will look so much better
;-)

Will do :-)
Kernel development is just my hobby so I'm not well versed in all of this.

You're doing great tbh!

quoted

quoted

quoted

quoted

  */

 /**

@@ -401,11 +416,14 @@ backoff:
 static int restore_fbdev_mode(struct drm_fb_helper *fb_helper)
 {
 	struct drm_device *dev = fb_helper->dev;
+	struct fb_info *info = fb_helper->fbdev;
 	struct drm_plane *plane;
 	int i;

 	drm_warn_on_modeset_not_all_locked(dev);

+	drm_fb_helper_dirty(info, 0, 0, info->var.xres, info->var.yres);

Why is this needed? If you do a modeset (or pageflip or whatever) drivers
are supposed to re-upload the entire screen. We've talked about adding a
dirty rectangle to atomic to allow userspace to optimize this, but there
should _never_ be a need to do a dirtyfb call around a modeset. Probably
just a driver bug in your panel drm drivers?

Ok, in tinydrm I now set a flag in &drm_simple_display_pipe_funcs
->plane_update to indicate that the next dirty() should do the whole
framebuffer which seems to work fine.
Should I actually perform the update as well?
If so I would need to add a worker in tinydrm to do that.

Yes, plane update should always do a full update. Not sure how you get
away with delaying that to ->dirty, maybe modesetting isn't
double-buffering when you don't have a GL that could do glamour.

->dirty is _only_ for frontbuffer rendering, not for page-flipping to an
entirely new buffer. In short if someone calls ->dirty on a buffer which
is currently not being displayed than a) they're silly b) drivers should
treat it as a no-op. Maybe we need a helper to do that ...
-Daniel

drm_fb_helper will call dirty() as long as there's fbdev activity, so the
driver needs to take that into account. For instance fbcon with a blinking
cursor will trigger calls even if a buffer has been set up on the drm side.
tinydrm checks the fb against the fb set on the plane and if it differs
it's a no-op.

Was really just an idea to make drivers a bit simpler, since pretty much
all of them we need to do this check. But with a grand total of just 3 (4
with tinydrm) implementing a non-trivial dirty callback that's not really
worth it I think.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch