Re: [RFC 1/6] drm: Add top level Kconfig option for DRM fbdev emulation

[Daniel Vetter <daniel@ffwll.ch>]

On Wed, Mar 25, 2015 at 01:47:54PM +0530, Archit Taneja wrote:
> Hi,
> 
> On 03/13/2015 02:36 PM, Daniel Vetter wrote:
> >On Fri, Mar 13, 2015 at 11:55:07AM +0530, Archit Taneja wrote:
> >>
> >>
> >>On 03/11/2015 08:47 PM, Daniel Vetter wrote:
> >>>On Wed, Mar 11, 2015 at 01:51:02PM +0530, Archit Taneja wrote:
> >>>>
> >>>>
> >>>>On 03/10/2015 05:47 PM, Daniel Vetter wrote:
> >>>>>On Tue, Mar 10, 2015 at 03:52:41PM +0530, Archit Taneja wrote:
> >>>>>>On 03/10/2015 03:35 PM, Daniel Vetter wrote:
> >>>>>>>On Tue, Mar 10, 2015 at 03:22:49PM +0530, Archit Taneja wrote:
> >>>>>>>>On 03/10/2015 03:17 PM, Daniel Vetter wrote:
> >>>>>>>>>On Tue, Mar 10, 2015 at 03:11:28PM +0530, Archit Taneja wrote:
> >>>>>>>>>>diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig
> >>>>>>>>>>index 151a050..38f83a0 100644
> >>>>>>>>>>--- a/drivers/gpu/drm/Kconfig
> >>>>>>>>>>+++ b/drivers/gpu/drm/Kconfig
> >>>>>>>>>>@@ -40,6 +40,24 @@ config DRM_KMS_FB_HELPER
> >>>>>>>>>>  	help
> >>>>>>>>>>  	  FBDEV helpers for KMS drivers.
> >>>>>>>>>>
> >>>>>>>>>>+config DRM_FBDEV_EMULATION
> >>>>>>>>>>+	bool "Enable legacy fbdev support for your modesetting driver"
> >>>>>>>>>>+	depends on DRM
> >>>>>>>>>>+	select DRM_KMS_HELPER
> >>>>>>>>>>+	select DRM_KMS_FB_HELPER
> >>>>>>>>>>+	select FB_SYS_FILLRECT
> >>>>>>>>>>+	select FB_SYS_COPYAREA
> >>>>>>>>>>+	select FB_SYS_IMAGEBLIT
> >>>>>>>>>>+	select FB_SYS_FOPS
> >>>>>>>>>>+	select FB_CFB_FILLRECT
> >>>>>>>>>>+	select FB_CFB_COPYAREA
> >>>>>>>>>>+	select FB_CFB_IMAGEBLIT
> >>>>>>>>>>+	default y
> >>>>>>>>>>+	help
> >>>>>>>>>>+	  Choose this option if you have a need for the legacy fbdev
> >>>>>>>>>>+	  support. Note that this support also provide the linux console
> >>>>>>>>>>+	  support on top of your modesetting driver.
> >>>>>>>>>
> >>>>>>>>>Maybe clarify that for linux console support you also need
> >>>>>>>>>CONFIG_FRAMEBUFFER_CONSOLE? fbdev alone isn't enough.
> >>>>>>>>
> >>>>>>>>DRM_KMS_FB_HELPER selects that for us, right?
> >>>>>>>
> >>>>>>>Hm right I've missed that. Reminds me that you need one more patch at the
> >>>>>>>end to remove all the various select DRM_KMS_FB_HELPER from all drm
> >>>>>>>drivers. Otherwise this knob here won't work by default if you e.g. select
> >>>>>>>radeon. In general we can't mix explicit options with menu entries with a
> >>>>>>>select.
> >>>>>>
> >>>>>>I was trying that out. Removing DRM_KMS_FB_HELPER and having
> >>>>>>DRM_FBDEV_EMULATION disabled breaks drivers which use FB stuff internally in
> >>>>>>their respective xyz_fbdev.c files.
> >>>>>
> >>>>>But with the stubbed out functions that should work, right? Why doesn't
> >>>>>it?
> >>>>
> >>>>There are still calls to functions from fb core like fb_set_suspend and
> >>>>register_framebuffer which aren't covered by the drm fb helper functions.
> >>>
> >>>Hm, sounds like we need another patch to stub out fb_set_suspend when
> >>>fbdev isn't enabled. Is there anything else?
> >>
> >>There are a handful of fb core functions which are called by drm drivers:
> >>
> >>fb_alloc_cmap/fb_dealloc_cmap
> >>
> >>fb_sys_read/fb_sys_write
> >>
> >>register_framebuffer/unregister_framebuffer/unlink_framebuffer/
> >>remove_conflicting_framebuffers
> >>
> >>fb_set_suspend
> >>
> >>fb_deferred_io_init/fb_deferred_io_cleanup
> >>
> >>framebuffer_alloc/framebuffer_release
> >
> >Hm yeah that's somewhat annoying indeed. What about the following:
> >1. We move all the #include <linux/fb.h> from drivers into drm_fb_helper.h
> >
> >2. Then we add stubs for these functions in drm_fb_helper.h, like this
> >
> >#if defined(CONFIG_FB)
> >#include <linux/fb.h>
> >#else
> >
> >/* static inline stubs for all the fb stuff used by kms drivers */
> >#endif
> >
> >Imo this makes sense since kms drivers really have a bit a special
> >situation with fbdev. They're not full-blown fbdev drivers and can be
> >useful fully without fbdev.
> >
> 
> I was trying this out. Removing 'linux/fb.h' and replacing stub fb funcs
> won't really work because struct declarations(like fb_info) also get
> removed.
> 
> I considered placing '#if IS_ENABLED(CONFIG_FB)' within linux/fb.h itself,
> but that seemed a bit too intrusive.
> 
> This is what I'm currently doing:
> 
> - Some funcs, like framebufer_alloc/release, alloc_cmap/dealloc_cmap would
> actually benefit if we have drm fb helpers for them. They are used in
> exactly the same manner by all the drivers.
> 
> - For the rest of the functions that are sparsely used, I was considering
> making very simple drm_fb_* wrapper functions. Something like:
> 
> void drm_fb_helper_deferred_io_init(struct drm_fb_helper *helper)
> {
> 	if (helper->fbdev)
> 		fb_deferred_io_init(helper->fbdev);
> }
> 
> We could have all fb calls called within drm_fb_helper.c, creating
> drm_fb_helper_* stub functions would then be an easier task. What do you
> think?

That's actually an option I considered, but I hoped we could do it with
less work. If this indeed works and you can create the patch that would be
awesome.

Thanks, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch