Re: [PATCH] of: treat PCI config space as IORESOURCE_MEM type

[Kumar Gala <galak@codeaurora.org>]

On May 29, 2014, at 8:41 PM, Liviu Dudau <liviu@dudau.co.uk> wrote:

> On Thu, May 29, 2014 at 07:29:31PM -0600, Bjorn Helgaas wrote:
>> On Thu, May 29, 2014 at 6:56 PM, Liviu Dudau <liviu@dudau.co.uk> wrote:
>>> On Thu, May 29, 2014 at 03:51:28PM -0500, Kumar Gala wrote:
>>>> 
>>>> On May 29, 2014, at 3:44 PM, Rob Herring <robherring2@gmail.com> wrote:
>>>> 
>>>>> On Thu, May 29, 2014 at 11:03 AM, Kumar Gala <galak@codeaurora.org> wrote:
>>>>>> If we have a PCI config space specified in something like a ranges
>>>>>> property we should treat it as memory type resource.
>>>>> 
>>>>> Config space should not be in ranges[1]. We have some cases that are,
>>>>> but we don't want new ones.
>>>> 
>>>> For the cases we have I agree, however an ECAM based cfg seems completely legit.
>>>> 
>>>>>> Signed-off-by: Kumar Gala <galak@codeaurora.org>
>>>>>> ---
>>>>>> drivers/of/address.c | 3 +++
>>>>>> 1 file changed, 3 insertions(+)
>>>>>> 
>>>>>> diff --git a/drivers/of/address.c b/drivers/of/address.c
>>>>>> index cb4242a..4e7ee59 100644
>>>>>> --- a/drivers/of/address.c
>>>>>> +++ b/drivers/of/address.c
>>>>>> @@ -122,6 +122,9 @@ static unsigned int of_bus_pci_get_flags(const __be32 *addr)
>>>>>>       u32 w = be32_to_cpup(addr);
>>>>>> 
>>>>>>       switch((w >> 24) & 0x03) {
>>>>>> +       case 0x00: /* cfg space */
>>>>>> +               flags |= IORESOURCE_MEM;
>>>>>> +               break;
>>>>> 
>>>>> How would you then distinguish actual memory ranges?
>>>> 
>>>> One assumes you are still looking at pci_space as part of of_pci_range
>>> 
>>> That doesn't happen when you start scanning the bus. The existing code will
>>> use the IORESOURCE_MEM for allocating memory space for devices, which is
>>> not what you want. Did you test your patch on any PCI system? I'm pretty
>>> sure that with my patch series that tries to make a generic framework for
>>> host controllers this will fail.
>>> 
>>> We really need a IORESOURCE_CFG flag for this space.
>> 
>> Maybe, but I'm not convinced yet.  The existing IORESOURCE_TYPE_BITS
>> types are for things that are mutually exclusive address spaces.  I
>> think this discussion is about ECAM, where the CPU side is definitely
>> in the same address space (IORESOURCE_MEM) as RAM, APICs, host bridge
>> apertures, device MMIO, etc.  The ECAM area must appear in the
>> iomem_resource tree so we avoid it when allocating other areas.
> 
> Agree, I'm only concerned that if this ECAM config space gets added to
> the list of pci_host_bridge windows it will be indistinguishable from
> IORESOURCE_MEM resources and pci_create_root_bus() will add it to the
> bus and allow devices present on that bus to be assigned addresses from
> that range. Which might not be what one wants for certain BARs.
> 
> I've had an aborted attempt to parse ECAM ranges in one version of my
> series (granted, I was trying to hack the IORESOURCE_TYPE_BITS as well)
> and things got horribly wrong quickly. I could give this patch a go with
> my series tomorrow when I'm in the office and report back.

We need to fix the parsing code to be smarter about this case.

- k
-- 
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation