On Wed, Jun 03, 2015 at 11:21:16AM +0100, Jiang Liu wrote:
> On 2015/6/3 18:03, Lorenzo Pieralisi wrote:
> > On Wed, Jun 03, 2015 at 10:36:19AM +0100, Jiang Liu wrote:
> >> On 2015/6/3 16:44, Hanjun Guo wrote:
> >>> On 2015???06???02??? 17:35, Lorenzo Pieralisi wrote:
> >>>> On Tue, Jun 02, 2015 at 07:12:53AM +0100, Jiang Liu wrote:
> >>>>> From: Hanjun Guo <hanjun.guo@xxxxxxxxxx>
> >>>>>
> >>>>> ARM64 ACPI based PCI host bridge init needs a arch dependent
> >>>>> struct pci_controller to accommodate common PCI host bridge
> >>>>> code which is introduced later, or it will lead to compile
> >>>>> errors on ARM64.
> >>>>>
> >>>>> Signed-off-by: Hanjun Guo <hanjun.guo@xxxxxxxxxx>
> >>>>> Tested-by: Suravee Suthikulpanit <Suravee.Suthikulpanit@xxxxxxx>
> >>>>> CC: Arnd Bergmann <arnd@xxxxxxxx>
> >>>>> CC: Catalin Marinas <catalin.marinas@xxxxxxx>
> >>>>> CC: Liviu Dudau <Liviu.Dudau@xxxxxxx>
> >>>>> CC: Lorenzo Pieralisi <Lorenzo.Pieralisi@xxxxxxx>
> >>>>> CC: Will Deacon <will.deacon@xxxxxxx>
> >>>>> Signed-off-by: Jiang Liu <jiang.liu@xxxxxxxxxxxxxxx>
> >>>>> ---
> >>>>> arch/arm64/include/asm/pci.h | 10 ++++++++++
> >>>>> 1 file changed, 10 insertions(+)
> >>>>>
> >>>>> diff --git a/arch/arm64/include/asm/pci.h b/arch/arm64/include/asm/pci.h
> >>>>> index b008a72f8bc0..70884957f253 100644
> >>>>> --- a/arch/arm64/include/asm/pci.h
> >>>>> +++ b/arch/arm64/include/asm/pci.h
> >>>>> @@ -10,6 +10,16 @@
> >>>>> #include <asm-generic/pci-bridge.h>
> >>>>> #include <asm-generic/pci-dma-compat.h>
> >>>>>
> >>>>> +struct acpi_device;
> >>>>> +
> >>>>> +struct pci_controller {
> >>>>> +#ifdef CONFIG_ACPI
> >>>>> + struct acpi_device *companion; /* ACPI companion device */
> >>>>> +#endif
> >>>>> + int segment; /* PCI domain */
> >>>>> + int node; /* NUMA node */
> >>>>> +};
> >>>>
> >>>> There is nothing ARM64 specific in this structure. The only
> >>>> reason I see you want to keep it arch specific is the iommu
> >>>> pointer on x86,
> >>>
> >>> And also plarform_data for IA64 too.
> >>>
> >>>> but I think we should find a way to make
> >>>> the common bits shared across archs (ie the struct above) and
> >>>> add (maybe a void*) to the generic struct to cater for arch
> >>>> specific data.
> >>>>
> >>>> Thoughts ?
> >>>
> >>> We discussed this already, it has limitations to make it
> >>> common to all archs, I think the limitation are:
> >>>
> >>> - struct pci_controller are also used for other archs
> >>> such as PowerPC and Tile, they will not use it for
> >>> ACPI purpose, so we can not used for all archs.
> >>>
> >>> - if we let struct pci_controller defined only for archs
> >>> using ACPI, such as introduce it in linux/acpi.h, we still
> >>> can not satisfy that the struct pci_controller is not
> >>> only used for ACPI case on x86, it will be used for
> >>> non-ACPI too.
> >>>
> >>> So it's pretty difficult to share it with across archs to me,
> >>> any more ideas?
> >> Hi Hanjun and Lorenzo,
> >> As mentioned by Hanjun, I have no idea yet about how to
> >> consolidating "struct pci_controller" further. One possible
> >> way is to move "struct pci_controller" related code into
> >> arch, but apparently that will reduce code reusing.
> >
> > I guess you can't move that struct pci_controller to generic code
> > since it is present on other archs too (with completely different
> > members).
> >
> > What you can do is creating a new struct (ie same purpose of pci_controller
> > with a different name) common to all archs that contains the common bits
> > + a void* data that contains arch specific data, and convert x86 and ia64
> > to using it.
> >
> > It is weird to be forced to declare a pci_controller structure in arm64
> > code with 0 arch specific data in it.
>
> Hi Lorenzo,
> I have thought to consolidate pci_controller for x86 and ia64,
> but that will make the change set much more bigger. How about to
> consolidate pci_controller by another patch set. That will be easier
> for review.
Agreed, but with this set you are forcing arm64 to define pci_controller
as pci_bus sysdata and I am not really keen on that, there are already
function calls in the arm64 pci layer that are there to make ACPI
compile and it is a bit annoying, instead of removing them we are
adding arch stuff on top.
How about passing a void* pointer (ie that is what pci_create_root_bus
expects) to acpi_pci_root_create through a member in acpi_pci_root_info
(I mean acpi_pci_root_info replaces the controller member with a void*
where you can add x86/ia64 pci_controller) ?
I understand this forces you to alloc the pci_controller in arch code,
but that's not a big deal right ? This way you can drop the
pci_controller struct from arm64 (I do not even know if it will ever
be needed, by looking at Hanjun's code, the bits of code that
need the pci_controller can be moved to generic PCI layer).
https://lkml.org/lkml/2015/5/26/215
This way we can add the generic struct we discussed later
(pci_controller refactoring), I agree it is going to be a bigger
change but at least you do not force something into arm64 that we do
not even know if it is required.
Thanks anyway for putting this series together.
Lorenzo
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
