[PATCH 05/11] x86, pci, acpi: Move arch-agnostic MMCONFIG (aka ECAM) and ACPI code out of arch/x86/ directory

[Lorenzo Pieralisi]

On Mon, Jun 08, 2015 at 03:57:38AM +0100, Hanjun Guo wrote:

[...]

> >>>>>> Why can't we make use of the ECAM implementation used by
> >>>>>> pci-host-generic
> >>>>>> and drivers/pci/access.c?
> >>>>>
> >>>>> We had that question when I had posted MMCFG patch set separately,
> >>>>> please see:
> >>>>> https://lkml.org/lkml/2015/3/11/492
> >>>>
> >>>> Yes, but the real question is, why do we need to have PCI config space
> >>>> up and running before a bus struct is even created ? I think the
> >>>> reason is
> >>>> the PCI configuration address space format (ACPI 6.0, Table 5-27, page
> >>>> 108):
> >>>>
> >>>> "PCI Configuration space addresses must be confined to devices on
> >>>> PCI Segment Group 0, bus 0. This restriction exists to accommodate
> >>>> access to fixed hardware prior to PCI bus enumeration".
> >>>>
> >>>> On HW reduced platforms I do not even think this is required at all,
> >>>> we have to look into this to avoid code duplication that might well
> >>>> turn out useless.
> >>>
> >>> This is only for the fixed hardware, which will be not available for
> >>> ARM64 (reduced hardware mode), but in Generic Hardware Programming
> >>> Model, we using OEM-provided ACPI Machine Language (AML) code to access
> >>> generic hardware registers, this will be available for reduced hardware
> >>> too.
> >>>
> >>> So in ACPI spec, it says: (ACPI 6.0 page 66, last paragraph)
> >>>
> >>> ACPI defines eight address spaces that may be accessed by generic
> >>> hardware implementations. These include:
> >>> * System I/O space
> >>> * System memory space
> >>> * PCI configuration space
> >>> * Embedded controller space
> >>> * System Management Bus (SMBus) space
> >>> * CMOS
> >>> * PCI BAR Target
> >>> * IPMI space
> >>>
> >>> So if any device using the PCI address space for control, such
> >>> as a system reset control device, its address space can be reside
> >>> in PCI configuration space (who can prevent a OEM do that crazy
> >>> thing? :) ), and it should be accessible before the PCI bus is
> >>> created.
> >>
> >> Us, by changing attitude and questioning features whose usefulness
> >> is questionable. I will look into this and raise the point, I am not
> >> thrilled by the idea of adding another set of PCI accessor functions
> >> and drivers because we have to access a register through PCI before
> >> enumerating the bus (and on arm64 this is totally useless since
> >> we are not meant to support fixed HW anyway). Maybe we can make acpica
> >> code use a "special" stub (ACPI specific, PCI configuration space address
> >> space has restrictions anyway), I have to review this set in its
> >> entirety to see how to do that (and I would kindly ask you to do
> >> it too, before saying it is not possible to implement it).
> >
> > I'm willing to do that, actually, if we don't need a mechanism to
> > access PCI config space before the bus is created, the code can be
> > simplified a lot.
> 
> After more investigation on the spec and the ACPI core code, I'm
> still not convinced that accessing to PCI config space before PCI
> bus creating is impossible, also there is no enough ARM64 hardware
> to prove that too. But I think we can go in this way, reuse the
> ECAM implementation by pci-host-generic for now, and implement the PCI
> accessor functions before enumerating PCI bus when needed in the
> future, does it make sense?

You mean we rewrite the patch to make sure we can use the PCI host generic
driver with MCFG and we leave the acpica PCI config call empty stubs on
arm64 (as they are now) ?

Thanks,
Lorenzo