On Wed, Jul 5, 2023, at 10:18, Thomas Zimmermann wrote:

Am 30.06.23 um 13:53 schrieb Arnd Bergmann:

quoted

On Fri, Jun 30, 2023, at 09:46, Thomas Zimmermann wrote:

quoted

Am 29.06.23 um 15:21 schrieb Arnd Bergmann:

I definitely get it for the screen_info, which needs the complexity.
For ARCHARCH_HAS_EDID_INFO I would hope that it's never selected by
anything other than x86, so I would still go with just a dependency
on x86 for simplicity, but I don't mind having the extra symbol if that
keeps it more consistent with how the screen_info is handled.

Well, I'd like to add edid_info to platforms with EFI. What would be 
arm/arm64 and loongarch, I guess. See below for the future plans.

To be clear: I don't mind using a 'struct edid_info' being passed
around between subsystems, that is clearly an improvement over
'struct screen_info'. It's the global variable that seems like
an artifact of linux-2.4 days, and I think we can do better than that.

quoted

quoted

quoted

I suppose you could use FIRMWARE_EDID on EFI or OF systems without
the need for a global edid_info structure, but that would not
share any code with the current fb_firmware_edid() function.

The current code is build on top of screen_info and edid_info. I'd
preferably not replace that, if possible.

One way I could imagine this looking in the end would be
something like

struct screen_info *fb_screen_info(struct device *dev)
{
       struct screen_info *si = NULL;

       if (IS_ENABLED(CONFIG_EFI))
             si = efi_get_screen_info(dev);

       if (IS_ENABLED(CONFIG_ARCH_HAS_SCREEN_INFO) && !si)
             si = screen_info;

       return si;
}

corresponding to fb_firmware_edid(). With this, any driver
that wants to access screen_info would call this function
instead of using the global pointer, plus either NULL pointer
check or a CONFIG_ARCH_HAS_SCREEN_INFO dependency.

This way we could completely eliminate the global screen_info
on arm64, riscv, and loongarch but still use the efi and
hyperv framebuffer/drm drivers.

If possible, I'd like to remove global screen_info and edid_info 
entirely from fbdev and the various consoles.

ok

We currently use screen_info to set up the generic framebuffer device in 
drivers/firmware/sysfb.c. I'd like to use edid_info here as well, so 
that the generic graphics drivers can get EDID information.

For the few fbdev drivers and consoles that require the global 
screen_info/edid_info, I'd rather provide lookup functions in sysfb 
(e.g., sysfb_get_screen_info(), sysfb_get_edid_info()). The global 
screen_info/edid_info state would then become an internal artifact of 
the sysfb code.

Hopefully that explains some of the decisions made in this patchset.

I spent some more time looking at the screen_info side, after my
first set of patches to refine the #ifdefs, and I think we don't
even need to make screen_info available to non-x86 drivers at all:

- All the vgacon users except for x86 can just register a static
  screen_info (or simplified into a simpler structure) with the
  driver itself. This even includes ia64, which does not support
  EFI framebuffers.

- The VESA, vga16, SIS, Intel and HyperV framebuffer drivers only
  need access to screen_info on x86. HyperV is the only driver that
  can currently access the data from EFI firmware on arm64, but
  that is only used for 'gen 1' guests, which I'm pretty sure
  only exist on x86.

- All the other references to screen_info are specific to EFI
  firmware, so we can move the global definition from arm,
  arm64, loongarch, riscv and ia64 into the EFI firmware
  code itself. It is still accessed by efifb and efi-earlycon
  at this point.

I have uploaded version 2 of my series to
https://git.kernel.org/pub/scm/linux/kernel/git/arnd/playground.git/log/?h=screen-info-v2
and will send it out after I get the green light from build
bots. 

       Arnd