Re: [PATCH] PCI: Only enable IO window if supported

[Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>]

On Wed, Jun 24, 2015 at 12:02:14AM +0100, Bjorn Helgaas wrote:
> On Tue, Jun 23, 2015 at 5:46 PM, Benjamin Herrenschmidt
> <benh@kernel.crashing.org> wrote:
> > On Tue, 2015-06-02 at 15:55 +0100, Lorenzo Pieralisi wrote:
> >> While at it, do you think it is reasonable to also claim the bridge
> >> windows (resources) in the respective pci_read_bridge_* calls ?
> >
> > No, don't claim in read. There's a clear distinction between gathering
> > resources and claiming them, and we need to keep that.
> >
> > Some fixups might happen in between the two for example.
> 
> Are there any existing fixups like that?  Concrete examples would help
> figure out the best way forward.
> 
> Most arches call pci_read_bridge_bases() from pcibios_fixup_bus().  I
> think that's a poor place to do it because it's code that normally
> doesn't have to be arch-specific.  Resource claiming is also usually
> done from arch code, and it shouldn't be arch-specific either.
> 
> If we move both the read and claim into generic code, then we might
> need to make sure there's a fixup phase in between or something.

Yes, that's where I am at the moment. On arm/arm64 PROBE_ONLY systems, if
I can't claim bridge apertures upon pci_read_bridge_bases, I can't
claim device resources in pcibios_add_device() since the bridge apertures
have not been claimed at that point, hence resulting in failures.

Given current code I see the following options:

(1) Claim bridge resources in pci_read_bridge_bases()
(2) Claim bridge resources in pcibios_add_device() (but that's horrible,
    since it requires looking up device upstream bridge and claim its
    resources)
(3) Do not claim resources on PROBE_ONLY systems (that's what arm does at
    present) and do not enable resources in pcibios_enable_device
(4) Add an initcall to arm/arm64 that carries out a resource survey,
    that's what's done on powerPC and also x86 it seems
    (eg pcibios_init in arch/powerpc/kernel/pci_64.c)

Personally I think (1) is by far the cleanest solution, I understand
Ben's concern but we need a way forward.

I will have to revert to (3) unless we find another solution, I would
like to make progress on this since it became a blocking issue.

As I said before, I will move pci_read_bridge_bases() to generic code
regardless.

Comments appreciated.

Thanks,
Lorenzo