Linux GOP initialization is wrong

[Hans de Goede]

Hi David,

On 4/28/20 1:58 AM, David Santamaría Rogado wrote:
> This is related to the issues at least on some devices for panel
> orientation quirks where added.

Thank you for looking into this.

> My tests have been done over a Lenovo ideapad D330.
> 
> This devices like the other ones that need panel orientation quirks,
> shows the initramfs with wrong stride and x and y swapped. By applying
> the panel orientation quirks this gets solved but many parts of the
> systems components needs to be patched. Hans has done a great job with
> plymouth, mutter... but always appears a new problem derived as for
> example vnc desktop sharing with this devices doesn't work and the
> output is send messed up.

When I first started adding support for devices which have their
screen mounted 90 degrees rotated my first attempts where aimed
at solving this transparently in the kernel.

Unfortunately this is not possible. On most affected devices
the hardware does not support 90 degrees rotation for the
primary display layer; or if it does this requires the framebuffer
being in a hardware-specific tiled format rather then being a
linear framebuffer. Using these tiled formats requires userspace
to be aware of this, which rules out transparently handling this
in the kernel.

Other layers (cursor layer, video overlay layers) have similar
issues which require userspace to be aware of what is happening,
so unfortunately there is no other way to deal with this then
fixing involved userspace components.

I'm a bit surprised that you sat that vnc desktop sharing does
not work, I guess that also depends on how the desktop sharing
works. If it pokes directly at the framebuffer somehow, then yes
it will be messed up. But if it goes through the display server
then things might work. I guess that it is possible that the
code doing this cannot deal with Xrandr output rotation ...

> The strange thing is that bootloaders like GRUB or rEFInd seems to be
> able to handle this and they paint themselves right, despite when
> booting Windows directly Windows paints itself right and booted with
> GRUB or rEFInd the first second also paint itself wrong. Haven't
> tested this too much but the interesting thing is in the next
> paragraph.

My experience with bootloaders showing themselves the right way
up is mixed. It seems that the firmware is doing some hacks for
this on some devices, at least for the EFI text console.

Funnily enough (for some definition of fun) on at least one of
the devices where the firmware is playing tricks (Asus T100HA IIRC)
the position of the carret for text-editing is off by one, which
is very annoying when editing the kernel commandline and which
clearly shows that things are being emulated in software here.

> I decided to get the UEFI GOP video modes and found that the D330 have
> these ones:
> Mode 0: 1200x1920
> Mode 1: 640x480
> Mode 2: 800x600
> Mode 3: 1024x768
> Mode 4: 1920x1200 (this is the default one started by the firmware)
> Mode 5: 480x640
> Mode 6: 600x800
> Mode 7: 768x1024
> 
> So I thought that Linux is taking the first mode despite is not the
> active one and that's why the display is messed up.

Nope, Linux does not touch the mode at all (nor does grub by default).
Doing a EFI/GOP modeset has the risk of triggering all sort of firmware
bugs. So we stick with what we get. This has interesting side effects
where on some systems you get a different mode when turning on the
machine and letting it boot, vs turning it on, pressing e.g. F12 to get
the boot menu and then boot Linux.

> Playing a little I could modify the GOP video mode before booting with
> the UEFI Shell by simple using the mode 150 101. This causes GOP video
> mode 5 to be switched to video mode 0, the first one. Booting now
> makes initramfs messages to be correctly rendered but in the wrong
> orientation.

Right, the rendering on the side thing is expected. As said the hardware
cannot do 90 degrees rotation with a linear framebuffer and the GOP
provided efifb is a linear framebuffer. So without telling the kernel
to software rotate its text console the text will always be on its side.

What your little EFI shell hack is doing is working around what seems to
be a bug on these Lenovo devices gives us the wrong stride and dimensions
for the EFI framebuffer.

Note that this very much is a Lenovo bug, all the other devices
with 90 degree rotated screens let us render the text console
on its side just fine. They correctly tells us the real size
and stride of the screen (so its portrait dimensions since it
is a portrait screen).

Even though this is a Lenovo bug we should probably still try to
find a way to deal with this though, so that the efifb works
correctly on these devices...

> A look at drivers/firmware/efi/libstub/gop.c seems to be what is
> happening, the first available video mode is used despite it could not
> be the active one in GOP and the active mode is not switched to the
> discovered one by Linux. Both GRUB and rEFInd are able to respect the
> video mode that GOP has active so it's possible to boot them landscape
> and portrait while being correctly rendered.
> 
> I think the video mode should not be the first discovered one but the
> active one, or at least, the highest resolution video mode that
> respects the orientation.

Again, Linux does not use the GOP concept of video modes at all,
it simply takes the active mode as reported by the UEFI and uses
that to show messages during early boot.

Also note that fixing the efifb is of little value early during
boot the kernel will load the i915 driver so that we can have
hardware rendered 3D, support for multiple monitors, etc. and as
soon as that is loaded the efifb settings no longer matter.

The i915 driver does not care about the GOP settings at all;
and without a quirk it too will cause everything to be rendered
on its side.

Regards,

Hans