Re: [PATCH] x86, ia64: Do not lose track of the EFI default VGA device

[Henrik Rydberg <rydberg@euromail.se>]

On 11/14/2014 03:42 PM, Bruno Prémont wrote:
> On Fri, 14 Nov 2014 13:53:30 +0100 Henrik Rydberg wrote:
>> Since commit 20cde694027e ("x86, ia64: Move EFI_FB
>> vga_default_device() initialization to pci_vga_fixup()") in the 3.17
>> merge window, the EFI framebuffer depends on the VGA arbitration
>> layer. However, the configuration does not reflect this, which leads
>> to a hard-to-find bug when FB_EFI is configured without VGA_ARB. Add a
>> select clause to remedy this.
> 
> Could you be more verbose in why it depends on/needs VGA_ARB?

When the EFI framebuffer is configured but not VGA_ARB, the kernel manages to
lose track of the default VGA device in some cases. As a result, the X11
nouveau driver fails on my MacBookAir3,1 (GeForce 320M, nv50, 0xaf), which
is booting in EFI_STUB mode.

The code to select the right PCI device was literally moved from efifb.c to the
internals of vgaarb. The PCI sysfs layer seems to depend on
vga_default_device(), which is only defined when VGA_ARB is set.

> With EFI starting to show up on ARM this is not necessarily true
> (no PCI -> no VGA_ARB arbitration).
> 
> So it would need to at least be select VGA_ARB if (PCI && !S390)
> in order to not have broken kernel configuration (in more or less
> exotic cases) while depends on VGA_ARB would be the only correct option
> if the rule 'select only allowed for leafs' is enforced.

I agree that the tie probably should be somewhere else. I am fine with any
combination of flags that respects the fact that efifb used to define
vga_default_device(), apparently for good reason.

Thanks,
Henrik