Re: [PATCH v3 2/3] pci: iproc: Add Broadcom iProc PCIe support

[Ray Jui <rjui@broadcom.com>]

Hi Dmitry,

On 3/6/2015 9:35 AM, Dmitry Torokhov wrote:
> On Fri, Mar 6, 2015 at 9:26 AM, Ray Jui <rjui@broadcom.com> wrote:
>> Hi Paul,
>>
>> On 3/6/2015 3:00 AM, Paul Bolle wrote:
>>> On Thu, 2015-03-05 at 17:01 -0800, Ray Jui wrote:
>>>> --- a/drivers/pci/host/Kconfig
>>>> +++ b/drivers/pci/host/Kconfig
>>>> @@ -106,4 +106,21 @@ config PCI_VERSATILE
>>>>      bool "ARM Versatile PB PCI controller"
>>>>      depends on ARCH_VERSATILE
>>>>
>>>> +config PCIE_IPROC
>>>> +    bool "Broadcom iProc PCIe controller"
>>>
>>> bool symbol.
>>>
>>>> +    help
>>>> +      This enables the iProc PCIe core controller support for Broadcom's
>>>> +      iProc family of SoCs. An appropriate bus interface driver also needs
>>>> +      to be enabled
>>>> +
>>>> +config PCIE_IPROC_PLTFM
>>>> +    bool "Broadcom iProc PCIe platform bus driver"
>>>
>>> Another bool symbol.
>>>
>>>> +    depends on ARCH_BCM_IPROC || COMPILE_TEST
>>>> +    depends on OF
>>>> +    select PCIE_IPROC
>>>> +    default ARCH_BCM_IPROC
>>>> +    help
>>>> +      Say Y here if you want to use the Broadcom iProc PCIe controller
>>>> +      through the generic platform bus interface
>>>> +
>>>>  endmenu
>>>
>>>> --- a/drivers/pci/host/Makefile
>>>> +++ b/drivers/pci/host/Makefile
>>>
>>>> +obj-$(CONFIG_PCIE_IPROC) += pcie-iproc.o
>>>> +obj-$(CONFIG_PCIE_IPROC_PLTFM) += pcie-iproc-pltfm.o
>>>
>>> Both objectfiles will never be part of a module.
>>>
>>>> --- /dev/null
>>>> +++ b/drivers/pci/host/pcie-iproc-pltfm.c
>>>
>>>> +#include <linux/module.h>
>>>
>>> Is this needed?
>>>
>>>> +MODULE_DEVICE_TABLE(of, iproc_pcie_of_match_table);
>>>
>>> This macro will be preprocessed away, won't it?
>>>
>>>> +static struct platform_driver iproc_pcie_pltfm_driver = {
>>>> +    .driver = {
>>>> +            .name = "iproc-pcie",
>>>> +            .of_match_table = of_match_ptr(iproc_pcie_of_match_table),
>>>> +            .suppress_bind_attrs = true,
>>>> +    },
>>>> +    .probe = iproc_pcie_pltfm_probe,
>>>> +};
>>>> +module_platform_driver(iproc_pcie_pltfm_driver);
>>>
>>> Perhaps it's clearer to have make this a call to
>>> platform_driver_register(), put that in a wrapper function, and invoke
>>> that wrapper function through device_initcall() or similar. I haven't
>>> actually tested that, so the details could be off.
>>>
>>>> +MODULE_AUTHOR("Ray Jui <rjui@broadcom.com>");
>>>> +MODULE_DESCRIPTION("Broadcom iPROC PCIe platform driver");
>>>> +MODULE_LICENSE("GPL v2");
>>>
>>> And these three macros will, effectively, be preprocessed away.
>>>
>>>> --- /dev/null
>>>> +++ b/drivers/pci/host/pcie-iproc.c
>>>
>>>> +#include <linux/module.h>
>>>
>>> See above.
>>>
>>>> +MODULE_AUTHOR("Ray Jui <rjui@broadcom.com>");
>>>> +MODULE_DESCRIPTION("Broadcom iPROC PCIe common driver");
>>>> +MODULE_LICENSE("GPL v2");
>>>
>>> Ditto.
>>>
>>>
>>> Paul Bolle
>>>
>>
>> So every single PCIe host driver under drivers/pci/host/* has their
>> config flag of "bool" type, and all of them except pci-keystone-dw.c
>> have these MODULE based macros in their driver. While I agree with you
>> that these macros will preprocessed away while compiled as statically
>> linked in, I thought it's a convention to have these macros in the
>> driver. At least it provides information on the author, driver
>> description, and license (although one can also argue you can find all
>> of those info from the maintainer list, Kconfig, and license header).
>>
>> Would you be able to sort this out with a developer or subsystem
>> maintainer who is familiar with this matter and let us know what is the
>> convention for MODULE macros usage in a driver when its config flag is
>> set to "bool" instead of "tristate"?
> 
> Is there a technical reason why the driver can't be a module? We can
> suppress module unloading if unloading properly is hard (i see you
> already suppress unbinding via sysfs), but we should be able to load
> it after kernel booted, no?
> 
> Thanks,
> Dmitry
> 

Although I have not tested it, but to my best knowledge there shouldn't
be any technical issue by making the PCIe iProc driver a loadable module
and installing the module later after the kernel finishes booting.

But I wonder why I haven't seen any PCIe host driver being allowed to be
compiled as module? Maybe there's no obvious benefit of doing that.
People typically opt to load slave devices as module but tend to keep
the bus or host devices compiled in.

And to allow a PCI host driver to be compiled as module, some PCI
functions need to have their symbols exported:

ERROR: "pci_common_swizzle" [drivers/pci/host/pcie-iproc.ko] undefined!
ERROR: "pci_fixup_irqs" [drivers/pci/host/pcie-iproc.ko] undefined!
ERROR: "pci_assign_unassigned_bus_resources"
[drivers/pci/host/pcie-iproc.ko] undefined!
ERROR: "pci_remove_root_bus" [drivers/pci/host/pcie-iproc.ko] undefined!
ERROR: "pci_stop_root_bus" [drivers/pci/host/pcie-iproc.ko] undefined!
ERROR: "pci_create_root_bus" [drivers/pci/host/pcie-iproc.ko] undefined!

Maybe Bjorn can help to shed some light here? Should I go ahead and
export_symbol these PCI functions and make the iProc PCI driver "tristate"?

Thanks,

Ray