RE: [PATCH v2 0/4] Allow non-legacy cards to be vgaarb default

Hi Daniel many thanks for your patch

-----Original Message-----
From: Daniel Axtens [mailto:dja@axtens.net]
Sent: 19 July 2017 03:15
To: linux-pci@vger.kernel.org; linuxppc-dev@lists.ozlabs.org; linux-
arm-kernel@lists.infradead.org
Cc: benh@kernel.crashing.org; Liuxinliang (Matthew Liu);
zourongrong@gmail.com; catalin.marinas@arm.com; will.deacon@arm.com;
Gabriele Paoloni; bhelgaas@google.com; airlied@linux.ie;
daniel.vetter@intel.com; alex.williamson@redhat.com; Daniel Axtens
Subject: [PATCH v2 0/4] Allow non-legacy cards to be vgaarb default
=20
[v2, in which I send the right patches. My apologies to you all.]
=20
Hi all,
=20
Previously I posted a patch that provided a quirk for a hibmc card
behind a particular Huawei bridge that allowed it to be marked as the
default device in the VGA arbiter.[0] This lead to some discussion.[1]
It was broadly suggested that a more generic solution would be better,
something in the style of powerpc's fixup_vga() quirk.
=20
Here is my suggested solution:
=20
 - Create a Kconfig option ARCH_WANT_VGA_ARB_FALLBACK and

In my opinion we could avoid depending on a Kernel config options.
I.e. we can have generic code that, after all PCI devs are enumerated:
1) check if we have a default vga device
2) if not check each registered PCI device and make default device the firs=
t
   one that is a VGA device, capable to respond to IO and MEM requests
   and that has a driver bound to it=20

=20
 - if an arch selects that option, install PCI_FIXUP_CLASS_ENABLE
   hook. This hook fires when a card is enabled, which will require
   that a driver has been bound.
=20
 - if there is no default device when the hook fires, and the device
   can control memory and I/O, mark it as default.

I am worried that the patchset you proposed has a race condition with the
VGA arbiter. In fact you see:
subsys_initcall(vga_arb_device_init) is not guaranteed to be called before
subsys_initcall(acpi_init)

acpi_init->acpi_scan_init->acpi_pci_root_init() at this stage the PCI enume=
ration
is done and as soon as a device is added the Kernel will look for a driver
to bind to it and therefore you quirk could be called before the VGA arbite=
r...
Do you agree?

What about modifying vgaarb.c to add a late_initcall() checking what I sugg=
ested
above?

Thanks
Gab

[...]