[PATCH] drivers/firmware: Move sysfb_init() from device_initcall to subsys_initcall_sync

[PATCH] drivers/firmware: Move sysfb_init() from device_initcall to subsys_initcall_sync

[Huacai Chen]

Consider a configuration like this:
1, efifb (or simpledrm) is built-in;
2, a native display driver (such as radeon) is also built-in.

Because efifb, radeon and sysfb are all in device_initcall() level, the
order in practise is like this:

efifb registered at first, but no "efi-framebuffer" device yet. radeon
registered later, and /dev/fb0 created. sysfb_init() comes at last, it
registers "efi-framebuffer" and then causes an error message "efifb: a
framebuffer is already registered". Make sysfb_init() to be subsys_
initcall_sync() can avoid this. But Javier Martinez Canillas has a more
general solution.

However, this patch still makes sense because it can make the screen
display as early as possible (We cannot move to subsys_initcall, since
sysfb_init() should be executed after PCI enumeration).

Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
---
 drivers/firmware/sysfb.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/firmware/sysfb.c b/drivers/firmware/sysfb.c
index 2bfbb05f7d89..aecf91517e54 100644
--- a/drivers/firmware/sysfb.c
+++ b/drivers/firmware/sysfb.c
@@ -80,4 +80,4 @@ static __init int sysfb_init(void)
 }
 
 /* must execute after PCI subsystem for EFI quirks */
-device_initcall(sysfb_init);
+subsys_initcall_sync(sysfb_init);
-- 
2.27.0

Re: [PATCH] drivers/firmware: Move sysfb_init() from device_initcall to subsys_initcall_sync

[Javier Martinez Canillas]

Hello Huacai,

Thanks for the patch.

On 5/21/22 15:41, Huacai Chen wrote:
> Consider a configuration like this:
> 1, efifb (or simpledrm) is built-in;
> 2, a native display driver (such as radeon) is also built-in.
>

As mentioned in the other thread, this is not a common configuration.

Usually the native display drivers are built as a module and only the
generic drivers that use a firmware-provided firmware for scanout are
built-in the kernel image.
 
> Because efifb, radeon and sysfb are all in device_initcall() level, the
> order in practise is like this:
> 
> efifb registered at first, but no "efi-framebuffer" device yet. radeon
> registered later, and /dev/fb0 created. sysfb_init() comes at last, it
> registers "efi-framebuffer" and then causes an error message "efifb: a
> framebuffer is already registered". Make sysfb_init() to be subsys_
> initcall_sync() can avoid this. But Javier Martinez Canillas has a more
> general solution.
>

You are talking about "[PATCH v5 0/7] Fix some races between sysfb device
registration and drivers probe", but unfortunately that approach was not
as general as needed, see:

https://lore.kernel.org/all/026b1c6d-c258-fa88-ed08-d1b5784c95b0@suse.de/

I need to go back to that series but I'm not sure when will have the time.
 
> However, this patch still makes sense because it can make the screen
> display as early as possible (We cannot move to subsys_initcall, since
> sysfb_init() should be executed after PCI enumeration).
>

Indeed. I agree that makes sense to move the platform device registration
early. As you mentioned this only solve some of the issues since we don't
know when the _driver_ is going to be registered. But at least we ensure
that by the time built-in drivers are registered, the platform device is
already there to match. Bringing the display output as early as possible.
 
> Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
> ---
>  drivers/firmware/sysfb.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/firmware/sysfb.c b/drivers/firmware/sysfb.c
> index 2bfbb05f7d89..aecf91517e54 100644
> --- a/drivers/firmware/sysfb.c
> +++ b/drivers/firmware/sysfb.c
> @@ -80,4 +80,4 @@ static __init int sysfb_init(void)
>  }
>  
>  /* must execute after PCI subsystem for EFI quirks */
> -device_initcall(sysfb_init);
> +subsys_initcall_sync(sysfb_init);

I would make the comment here more verbose. Mentioning that PCI enumeration
happens in the subsys_initcall() init call level and so is safe to register
this at subsys_initcall_sync(), which happens after or something like that.

Probably would be good to also mention the same in the patch description.

Reviewed-by: Javier Martinez Canillas <javierm@redhat.com>

-- 
Best regards,

Javier Martinez Canillas
Linux Engineering
Red Hat